Coordinate Attention for Efficient Mobile Network Design

🎯 Pitch

Standard channel attention (like SE blocks) throws away positional information by using 2D global pooling—a surprisingly costly omission for mobile networks, where coordinate attention recovers it by factorizing pooling into two 1D feature encodings along horizontal and vertical directions. This simple change yields dramatic gains on dense prediction tasks (e.g., +2.5 mIoU on segmentation and +2.2 AP on detection over SE) while adding negligible compute, effectively making mobile networks spatially aware for free.

1. Executive Summary

This paper proposes coordinate attention, a novel light-weight attention mechanism for mobile networks that embeds positional information into channel attention by factorizing 2D global pooling into two parallel 1D feature encoding processes—aggregating features along the horizontal and vertical directions separately to produce direction-aware and position-sensitive attention maps (as opposed to SE attention's single channel descriptor or CBAM's local 7×7 convolution). On ImageNet classification with MobileNetV2, coordinate attention yields a 0.8% top-1 accuracy gain over the SE block (reaching 74.3% vs. 73.5%) with comparable parameters and computational cost, while on downstream dense prediction tasks the advantage widens substantially—improving semantic segmentation mIoU by 2.5 percentage points on Pascal VOC 2012 (73.32% vs. 70.84% at output stride 16) and COCO object detection AP by 2.2 points (24.5 vs. 22.3), establishing that positional information encoding through 1D directional feature aggregation benefits vision tasks most when the output space is dense rather than scalar.

2. Context and Motivation

The Core Problem: Mobile Networks Sacrifice Spatial Information for Efficiency

The fundamental problem this paper addresses is that mobile neural networks lack an efficient mechanism to encode precise positional information while maintaining the computational constraints required for deployment on resource-limited devices. This matters because mobile networks—networks designed to run with limited parameters, memory, and compute on smartphones, embedded systems, and edge devices—are fundamentally constrained by what they can see spatially. A standard convolution operator captures local spatial relationships within its kernel window, but modeling long-range spatial dependencies (knowing that two distant pixels belong to the same object, or understanding where in the image an object exists) typically requires mechanisms that are too expensive for mobile budgets.

This gap creates a tension that the paper frames explicitly: attention mechanisms have been proven extraordinarily effective at improving model performance by telling the network "what" and "where" to attend (Section 1), but the most powerful attention designs—self-attention, non-local blocks, and sophisticated spatial attention modules—carry computational overhead that makes them "not affordable for mobile networks" (Section 1). The consequence is that mobile network design has settled for attention mechanisms that are computationally cheap but spatially impoverished, leaving potentially significant accuracy gains on the table, particularly for tasks where knowing where objects are located is critical.

Why This Matters: The Gap Between Classification and Dense Prediction

The practical significance of this problem becomes stark when comparing classification performance against downstream dense prediction tasks. The paper's motivating observation (visible in Figure 1 and quantified in Section 4) is that existing attention methods like the Squeeze-and-Excitation (SE) block provide reasonable gains on ImageNet classification—a task where the network only needs to produce a single scalar label per image—but their benefits are much more limited for object detection and semantic segmentation, where the network must make predictions at every spatial location.

This discrepancy reveals a structural weakness: classification can succeed with attention mechanisms that only model which channels are important (channel attention), because the spatial "where" information gets collapsed into a global average that is sufficient for distinguishing "dog" from "cat" but not for drawing a precise boundary around the dog. Dense prediction tasks, by contrast, demand that the network maintain and refine spatial information throughout its layers. An attention mechanism that discards positional information—either by collapsing the entire spatial dimension into a scalar (SE) or by capturing only local spatial context through a single convolution layer (CBAM)—leaves the network starved of the long-range spatial reasoning needed to, for example, segment a large object spanning much of the image or detect small objects at precise coordinates.

The paper makes this connection explicit in Section 4.4 when discussing why coordinate attention shows larger gains on segmentation than on classification:

"We argue that this is because our coordinate attention is able to capture long-range dependencies with precise positional information, which is more beneficial to vision tasks with dense predictions, such as semantic segmentation."

This observation positions the problem as not merely academic—getting mobile networks to work well on dense prediction tasks has direct real-world impact for applications like autonomous driving (where Cityscapes segmentation matters), augmented reality (where object detection on mobile devices is core), and mobile photography (where semantic segmentation enables portrait mode and background effects).

Prior Approaches and Where They Fall Short

The paper identifies three lines of prior work that address pieces of the problem but each leave crucial gaps.

Squeeze-and-Excitation (SE) attention (Hu et al., 2018). The SE block, introduced in SENet and illustrated in Figure 2(a), is the dominant attention mechanism in mobile networks. Its operation can be understood in three stages: (1) squeeze the spatial dimensions of the input feature map using global average pooling, producing a single scalar per channel representing the channel's global activation; (2) excitation through a small bottleneck network (two fully-connected layers with a non-linearity) that learns to predict channel-wise importance weights; and (3) recalibration by multiplying the original feature map channel-wise by these learned weights. The mathematical formulation (Section 3.1, Eqn. 1) compresses each channel $c$ of spatial size $H \times W$ into a single value $z_c$ :

$z_c = \frac{1}{H \times W} \sum_{i=1}^{H} \sum_{j=1}^{W} x_c(i, j)$

The SE block's strengths are clear: it adds negligible parameters (roughly $2C^2/r$ for a reduction ratio $r$ , typically $r=16$ or $24$ ) and computation, it models inter-channel dependencies explicitly, and it provides consistent gains across architectures. However, its fundamental limitation is equally clear from Eqn. 1: by averaging across the entire spatial extent, the squeeze operation annihilates all positional information. The resulting channel descriptor $z_c$ knows whether a channel was active somewhere in the image, but has no knowledge of where. As the paper states:

"It only considers reweighing the importance of each channel by modeling channel relationships but neglects positional information, which as we will prove experimentally in Section 4 to be important for generating spatially selective attention maps." (Section 3.1)

This is not a minor omission—spatial structure is the defining characteristic of visual data, and discarding it limits what the attention mechanism can learn to emphasize. An SE block can learn "this channel detects edges, so weight it highly" but cannot learn "this channel detects edges on the left side, which is where the object of interest tends to appear."

CBAM (Convolutional Block Attention Module, Woo et al., 2018). CBAM, shown in Figure 2(b), attempts to remedy the spatial blindness of SE attention by adding a spatial attention module in sequence after the channel attention. The spatial attention module operates by: (1) aggregating channel information through average pooling and max pooling across the channel dimension, producing a 2-channel spatial descriptor; (2) applying a $7 \times 7$ convolution to this descriptor to capture local spatial context; and (3) generating a single-channel spatial attention map that weights each spatial location.

This design explicitly reintroduces spatial information, but the paper identifies two specific weaknesses that prevent it from being a satisfying solution for mobile networks. First, channel squeezing causes information loss: compressing the entire channel dimension (which could be hundreds of channels in intermediate layers) down to just 2 channels (avg-pool and max-pool) discards substantial information about which feature types are active at which locations. The paper states:

"The spatial attention module in CBAM squeezes the channel dimension to 1, leading to information loss. However, our coordinate attention uses an appropriate reduction ratio to reduce the channel dimension in the bottleneck, avoiding too much information loss." (Section 4.3)

Second, the $7 \times 7$ convolution only captures local spatial relationships. A single convolution layer with a fixed kernel size has a limited receptive field—it can relate pixels within a $7 \times 7$ neighborhood but cannot model dependencies between distant spatial locations. The paper argues this is a fundamental limitation:

"CBAM utilizes a convolutional layer with kernel size $7 \times 7$ to encode local spatial information while our coordinate attention encodes global information by using two complementary 1D global pooling operations. This enables our coordinate attention to capture long-range dependencies among spatial locations that are essential for vision tasks." (Section 4.3)

The distinction between local and global spatial encoding is central to the paper's argument. Vision tasks routinely involve objects that span large portions of an image, requiring the network to relate features at distant spatial coordinates—for example, determining whether two separated edges belong to the same object contour. A $7 \times 7$ convolution cannot do this without being stacked many times (which would defeat the purpose of a lightweight attention module), while a global pooling operation captures long-range context in a single step.

Non-local/self-attention networks (NLNet, GCNet, A²Net, et al.). A third class of attention mechanisms, based on non-local operations or self-attention, explicitly computes pairwise relationships between all spatial positions. These methods can capture truly global spatial dependencies—every position can attend to every other position. However, the computational cost is prohibitive for mobile networks. A standard non-local block computes an attention matrix of size $HW \times HW$ , where $H$ and $W$ are the spatial dimensions, resulting in quadratic complexity in the number of spatial positions. For a feature map of size $14 \times 14$ (a typical intermediate resolution in mobile networks), this is $196 \times 196 = 38,416$ pairwise scores—tractable perhaps in a large server-side model but too expensive per inference for a mobile device where every millisecond and milliwatt counts.

The paper acknowledges this body of work (Section 2.2) but positions it as fundamentally misaligned with mobile constraints:

"Non-local/self-attention networks are recently very popular due to their capability of building spatial or channel-wise attention... However, because of the large amount of computation inside the self-attention modules, they are often adopted in large models but not suitable for mobile networks."

This creates a clear design space: SE attention is cheap but spatially blind, non-local attention is spatially rich but expensive, and CBAM attempts a middle ground but is limited by local spatial encoding and channel squeezing. The gap—an attention mechanism that is simultaneously (1) computationally efficient, (2) capable of encoding long-range spatial dependencies, and (3) able to preserve precise positional information—is the opportunity this paper seizes.

How This Paper Positions Itself

The paper introduces coordinate attention as a mechanism that sits at a deliberately chosen point in the design space between SE's channel-only encoding and full non-local spatial attention. The core insight that enables this positioning is the factorization of 2D spatial encoding into two orthogonal 1D encoding processes.

Rather than collapsing both spatial dimensions simultaneously (SE's 2D global pooling) or attempting to relate all pairs of spatial positions (non-local's $HW \times HW$ attention), coordinate attention decomposes spatial information into two independent streams: one that aggregates features horizontally (producing a $C \times H \times 1$ feature map that knows which rows contain important features) and one that aggregates vertically (producing a $C \times 1 \times W$ feature map that knows which columns contain important features). The paper argues this factorization achieves a specific and valuable property:

"These two transformations also allow our attention block to capture long-range dependencies along one spatial direction and preserve precise positional information along the other spatial direction, which helps the networks more accurately locate the objects of interest." (Section 3.2.1)

This is a subtle but powerful claim. By pooling along only one dimension at a time, the resulting feature map retains spatial resolution along the orthogonal dimension. A horizontal pooling operation (Eqn. 4) produces an output at each vertical position $h$ , meaning the network knows the row where features are active. A vertical pooling operation (Eqn. 5) produces an output at each horizontal position $w$ , meaning the network knows the column where features are active. When these two attention maps are multiplied element-wise onto the original feature map (Eqn. 9), the network effectively attends to specific $(i,j)$ coordinate regions where both the row-attention and column-attention signals are strong—a form of coordinate-aware spatial attention without ever explicitly computing a full $H \times W$ attention matrix.

The paper explicitly contrasts this with both predecessors. Against SE attention: unlike SE's single scalar per channel, coordinate attention produces two 1D attention vectors per channel, preserving positional information that SE discards. Against CBAM: unlike CBAM's local $7 \times 7$ convolution, coordinate attention's 1D global pooling operations each have a receptive field spanning the entire image along their respective dimension (width-wise for vertical pooling, height-wise for horizontal pooling), enabling long-range dependency capture. Against non-local attention: unlike the quadratic cost of pairwise self-attention, coordinate attention's cost scales as $O(C \times (H+W))$ —linear in the spatial dimensions—making it viable for mobile deployment.

The paper also positions itself within the broader mobile network architecture landscape by demonstrating integration into three distinct backbone families: MobileNetV2's inverted residual blocks, MobileNeXt's sandglass bottleneck blocks, and EfficientNet's NAS-derived architecture. This is not merely a demonstration of versatility—it establishes that coordinate attention is a drop-in replacement for existing attention mechanisms, not a method that requires custom architecture design. The integration diagrams in Figure 3 show attention blocks inserted at identical positions where SE or CBAM would be placed, with the same residual connection patterns, reinforcing this message of compatibility.

Finally, the paper's positioning in relation to downstream tasks is a deliberate rhetorical and experimental choice. Rather than evaluating only on ImageNet classification (where attention gains often saturate or are modest), the paper emphasizes dense prediction tasks—COCO object detection, Pascal VOC and Cityscapes semantic segmentation—where spatial information is most critical and where the limitations of spatially-blind attention should be most apparent. This choice of evaluation suite is itself an argument: if coordinate attention's claimed advantage is better spatial encoding, then the strongest evidence should appear in tasks where spatial precision matters most. The paper's results bear this out, with segmentation improvements (+2.5 mIoU on VOC) substantially exceeding classification improvements (+0.8% top-1 on ImageNet).

3. Technical Approach

3.1 Reader Orientation

The paper develops a lightweight computational module—the coordinate attention block—that can be inserted into existing mobile neural network architectures to make them more sensitive to where objects are located in an image, not just what features are present. The core problem it solves is that existing efficient attention mechanisms for mobile networks either discard all positional information (Squeeze-and-Excitation) or capture only local positional relationships at the cost of channel information (CBAM), leaving the network unable to efficiently model long-range spatial dependencies—knowing that two distant pixels belong to the same object, or precisely locating an object within the image—without incurring the quadratic computational cost of full self-attention.

3.2 Big-Picture Architecture (Diagram in Words)

The coordinate attention block takes a 3D feature tensor as input (channels × height × width) and outputs a tensor of identical dimensions, where each element has been reweighted by learned attention weights that encode which rows and columns contain objects of interest. The system has four sequential stages:

Coordinate Information Embedding — Two parallel 1D global pooling operations: one pools horizontally (producing a $C \times H \times 1$ tensor encoding per-row information) and one pools vertically (producing a $C \times 1 \times W$ tensor encoding per-column information). This factorizes the 2D spatial collapse that SE attention performs into two orthogonal 1D encodings that preserve spatial structure along each direction.
Shared Feature Transformation — The two direction-aware feature maps are concatenated along the spatial dimension and passed through a shared $1 \times 1$ convolution that reduces the channel dimension by a factor $r$ , producing a compact intermediate representation $f \in \mathbb{R}^{C/r \times (H+W)}$ that encodes spatial information from both directions in a single bottleneck.
Split and Independent Encoding — The intermediate representation is split back into two separate tensors (one for height, one for width), which are independently transformed via two separate $1 \times 1$ convolutions back to the original channel dimension and passed through a sigmoid activation to produce attention weights in the range $[0, 1]$ .
Coordinate-Aware Recalibration — The two 1D attention weight vectors ( $g^h$ for height and $g^w$ for width) are expanded via broadcasting and multiplied element-wise onto the original input feature map, such that each spatial position $(i,j)$ is reweighted by the product of the attention weight for row $i$ and the attention weight for column $j$ .

The entire block is wrapped in a residual connection: the output is added to the input (or, in the typical mobile network integration pattern, the attention block is inserted at the position where SE attention would go—after the depthwise convolution in an inverted residual block or sandglass block).

3.3 Roadmap for the Deep Dive

First, the formal definition of SE attention (Section 3.1) as the baseline that coordinate attention builds upon, establishing why 2D global pooling loses positional information and setting up the mathematical contrast.
Second, the coordinate information embedding stage (Section 3.2.1) — the two 1D pooling operations, their mathematical formulation, and the crucial property they provide: capturing long-range dependencies along one direction while preserving precise positional information along the orthogonal direction.
Third, the coordinate attention generation stage (Section 3.2.2) — how the concatenated direction-aware features are transformed, split, and independently encoded into attention weights, including the design rationale for the shared $1 \times 1$ convolution and the reduction ratio $r$ .
Fourth, the final recalibration operation (Eqn. 9) and how the two 1D attention maps combine multiplicatively to produce spatially selective attention without ever computing a full $H \times W$ attention matrix.
Fifth, the integration into mobile network architectures (Section 3.3) — exactly where the attention block is inserted in inverted residual blocks and sandglass blocks, and why these insertion points are chosen.

3.4 Detailed, Sentence-Based Technical Breakdown

This is primarily a method paper whose core idea is that factorizing 2D spatial pooling into two orthogonal 1D pooling operations enables a mobile-efficient attention mechanism that simultaneously encodes channel relationships, long-range spatial dependencies, and precise positional information—three properties that no prior lightweight attention mechanism provided together.

Squeeze-and-Excitation Attention as the Baseline to Improve Upon

The paper builds on SE attention not as a competitor to be outperformed but as a foundation to be extended. Understanding SE's mathematical formulation is essential because coordinate attention's innovations are defined in direct contrast to SE's limitations.

The Squeeze Operation (Global Information Embedding). Given an input feature tensor $X = [x_1, x_2, \ldots, x_C] \in \mathbb{R}^{C \times H \times W}$ (where $C$ is the number of channels, $H$ is height, and $W$ is width), the squeeze step compresses the entire spatial extent of each channel into a single scalar by global average pooling:

$z_c = \frac{1}{H \times W} \sum_{i=1}^{H} \sum_{j=1}^{W} x_c(i, j)$

where $x_c(i, j)$ is the activation value at spatial position $(i, j)$ in channel $c$ , and $z_c \in \mathbb{R}$ is the resulting scalar descriptor for that channel. This produces a vector $z \in \mathbb{R}^C$ —one number per channel, representing that channel's average activation across the entire image.

What it computes: For each channel, the operation sums all $H \times W$ spatial activation values and divides by the total number of spatial positions, collapsing a 2D feature map into a single global average. The output is a $C$ -dimensional vector that encodes how active each channel is on average, but with zero spatial information remaining.

Why this form and its limitation: Global average pooling is computationally trivial (it adds essentially no parameters and negligible FLOPs) and provides a summary statistic that captures whether a feature detector fired somewhere in the image. This is sufficient for distinguishing object classes in image classification—if a "cat ear detector" channel has high average activation, the image likely contains a cat, regardless of where the ears are. However, as the paper notes, this operation "squeezes global spatial information into a channel descriptor and hence is difficult to preserve positional information." The mathematical operation is a surjective (many-to-one) mapping from $\mathbb{R}^{H \times W}$ to $\mathbb{R}$ ; every spatial arrangement that produces the same average produces the same squeeze output, making position irrecoverable.

The Excitation Operation (Adaptive Channel Recalibration). The vector $z$ is passed through a bottleneck network consisting of two fully-connected layers with a non-linear activation between them, followed by a sigmoid to produce channel-wise attention weights in the range $(0, 1)$ :

$\hat{z} = T_2(\text{ReLU}(T_1(z)))$

where $T_1$ is a linear transformation that reduces the dimensionality from $C$ to $C/r$ , and $T_2$ is a linear transformation that restores it to $C$ . The reduction ratio $r$ controls the bottleneck size—a smaller $r$ means more parameters in the bottleneck (the paper uses $r=24$ for default experiments and tests $r=16$ and $r=12$ in ablations).

What it computes: A small multi-layer perceptron learns to predict, from the global channel statistics $z$ , a vector of importance weights $\hat{z} \in (0,1)^C$ that tell the network which channels to emphasize and which to suppress. The bottleneck structure forces the network to learn a compact encoding of channel interdependencies, acting as a regularizer and parameter-efficient design.

The Recalibration Step: The final output of the SE block is:

$\hat{X} = X \cdot \sigma(\hat{z})$

where $\sigma$ is the sigmoid function, and the multiplication is channel-wise: $\hat{x}_c = x_c \cdot \sigma(\hat{z}_c)$ . Each channel's feature map is uniformly scaled by a single scalar weight. Important channels (those the excitation network predicts should contribute strongly) are preserved or amplified; unimportant channels are suppressed.

Why SE is insufficient for spatial encoding: The recalibration applies the same weight to every spatial position within a channel. If channel $c$ has weight $0.8$ , all $H \times W$ positions in that channel's feature map are multiplied by $0.8$ equally. The attention mechanism cannot make channel $c$ more influential on the left side of the image and less influential on the right—it has no spatial resolution at all. This is the fundamental limitation that coordinate attention addresses.

Coordinate Information Embedding: Factorizing 2D Pooling into Two 1D Encodings

The paper's core architectural innovation is replacing the single 2D global pooling operation in SE attention with two parallel 1D pooling operations. The motivation is stated directly:

"To encourage attention blocks to capture long-range interactions spatially with precise positional information, we factorize the global pooling as formulated in Eqn. (1) into a pair of 1D feature encoding operations." (Section 3.2.1)

Horizontal (X-direction) pooling. A pooling kernel of spatial extent $(H, 1)$ operates along the vertical direction: for each channel and each row (horizontal line of pixels), the operation averages all activation values across the width dimension:

$z_c^h(h) = \frac{1}{W} \sum_{0 \leq i < W} x_c(h, i)$

where $x_c(h, i)$ is the activation of channel $c$ at row $h$ and column $i$ , and $z_c^h(h)$ is the resulting value for channel $c$ at vertical position $h$ . The output is a tensor $z^h \in \mathbb{R}^{C \times H \times 1}$ —for each of the $C$ channels, there is one value per row of the original feature map.

What it computes: For each channel and each row independently, average all pixel values across that entire row. If the original feature map is $C \times H \times W$ , the result is $C \times H \times 1$ . Each element $z_c^h(h)$ tells the network how strongly channel $c$ activates, on average, along the horizontal strip at vertical position $h$ .

Why this preserves positional information along one direction: Unlike SE's 2D pooling which collapses both $H$ and $W$ into a single number, this operation preserves the vertical coordinate $h$ . The output has $H$ distinct values per channel, each corresponding to a specific row. The network knows which row produced each activation value. Long-range dependencies are captured along the horizontal direction (because summing across all columns $W$ relates all positions in a row), while precise positional information is preserved along the vertical direction (because the output retains separate values for each $h$ ).

Vertical (Y-direction) pooling. Symmetrically, a pooling kernel of spatial extent $(1, W)$ operates along the horizontal direction: for each channel and each column, average all activation values across the height dimension:

$z_c^w(w) = \frac{1}{H} \sum_{0 \leq j < H} x_c(j, w)$

where $x_c(j, w)$ is the activation of channel $c$ at row $j$ and column $w$ , and $z_c^w(w)$ is the resulting value for channel $c$ at horizontal position $w$ . The output is a tensor $z^w \in \mathbb{R}^{C \times 1 \times W}$ —for each channel, one value per column.

What it computes: For each channel and each column independently, average all pixel values across that entire column. Each element $z_c^w(w)$ tells the network how strongly channel $c$ activates, on average, along the vertical strip at horizontal position $w$ .

The complementary property produced by the pair: The horizontal pooling captures long-range dependencies along the width dimension (by averaging across all columns) while preserving precise vertical position (by maintaining separate outputs per row). The vertical pooling captures long-range dependencies along the height dimension (by averaging across all rows) while preserving precise horizontal position (by maintaining separate outputs per column). When these two encodings are combined multiplicatively at the final recalibration step, the resulting attention mechanism can distinguish specific $(h, w)$ coordinate regions: a region where both $g^h(h)$ and $g^w(w)$ are high will be strongly attended, while regions where either attention value is low will be suppressed. The paper explicitly claims this structure "helps the networks more accurately locate the objects of interest."

Computational cost of this factorization. Each 1D pooling operation requires summing across one full spatial dimension. A horizontal pooling requires $O(C \times H \times W)$ additions (one addition per element of the input tensor, since each pixel contributes to one horizontal sum) and produces $C \times H$ output values. Similarly, a vertical pooling requires $O(C \times H \times W)$ additions and produces $C \times W$ output values. The total cost of both pooling operations together is $O(C \times H \times W)$ —exactly linear in the size of the input tensor—which is the same asymptotic cost as the single 2D global pooling in SE attention (which also sums over all $H \times W$ elements per channel). The factorization does not increase the computational order, merely reorganizes the output shape from $C \times 1 \times 1$ to $C \times (H + W) \times 1$ .

Coordinate Attention Generation: From Direction-Aware Features to Attention Weights

With the two direction-aware feature maps $z^h \in \mathbb{R}^{C \times H \times 1}$ and $z^w \in \mathbb{R}^{C \times 1 \times W}$ produced by the coordinate information embedding stage, the coordinate attention generation stage transforms these into attention weight tensors that can be applied to the original input. The paper lays out three explicit design criteria for this transformation (Section 3.2.2): it must be "as simple and cheap as possible" for mobile deployment; it must "make full use of the captured positional information"; and it must "effectively capture inter-channel relationships."

Concatenation and shared $1 \times 1$ convolution. The first step concatenates the two direction-aware feature maps along the spatial dimension, aligning them before the shared transformation:

$f = \delta(F_1([z^h, z^w]))$

where $[\cdot, \cdot]$ denotes concatenation, $F_1$ is a shared $1 \times 1$ convolutional transformation, $\delta$ is a non-linear activation function (the paper uses this generic notation but in practice the activation is typically ReLU, consistent with standard mobile network designs), and $f \in \mathbb{R}^{C/r \times (H+W)}$ is the intermediate feature map encoding spatial information from both directions.

What happens in detail: The tensor $z^h$ of shape $C \times H \times 1$ and the tensor $z^w$ of shape $C \times 1 \times W$ need to be combined in a way that allows the network to learn relationships between the vertical and horizontal encodings. Concatenation along the spatial dimension produces a tensor of shape $C \times 1 \times (H+W)$ (or equivalently $C \times (H+W)$ after squeezing the singleton dimension). A $1 \times 1$ convolution then operates on each of the $(H+W)$ spatial positions independently, applying the same linear transformation followed by non-linearity, producing a tensor of shape $C/r \times (H+W)$ .

Why a shared $1 \times 1$ convolution: The $1 \times 1$ convolution has three important properties. First, it is parameter-efficient: a $1 \times 1$ convolution from $C$ channels to $C/r$ channels uses $C \times C/r + C/r$ parameters (for the non-existent "spatial" dimension, since $1 \times 1$ convolutions are just per-position linear transformations), which is exactly the same cost as one fully-connected layer in SE attention's excitation bottleneck. Second, by sharing the same transformation weights across all $(H+W)$ spatial positions, it forces the network to learn a common encoding of spatial-direction information that applies equally to horizontal and vertical position encodings—this parameter sharing acts as a regularizer and reduces overfitting. Third, applying it to the concatenated tensor allows the transformation to mix information from the horizontal and vertical encodings at corresponding spatial positions before they are split and processed independently.

Why this avoids the channel squeezing problem of CBAM: CBAM's spatial attention module reduces the entire channel dimension to 2 (average-pooled and max-pooled values per spatial position), then applies a $7 \times 7$ convolution. This means the spatial attention map is computed from only 2 channels of information, losing the rich channel-specific information about which features are active at each position. Coordinate attention, by contrast, uses a $1 \times 1$ convolution that reduces the channel dimension to $C/r$ —with $C$ typically being 64–1280 and $r$ being 16–32, this means the intermediate representation $f$ retains tens to hundreds of channels of information per spatial position, rather than just 2. The paper states this explicitly:

"CBAM squeezes the channel dimension to 1, leading to information loss. However, our coordinate attention uses an appropriate reduction ratio to reduce the channel dimension in the bottleneck, avoiding too much information loss." (Section 4.3)

Splitting and independent encoding. The intermediate feature map $f$ is split along the spatial dimension back into two separate tensors corresponding to the original horizontal and vertical encodings:

$f^h \in \mathbb{R}^{C/r \times H}, \quad f^w \in \mathbb{R}^{C/r \times W}$

Each of these is then passed through its own $1 \times 1$ convolutional transformation (effectively, a separate per-position linear layer) to project back to the original channel dimension $C$ , followed by a sigmoid activation to produce attention weights:

$g^h = \sigma(F_h(f^h))$

$g^w = \sigma(F_w(f^w))$

where $F_h$ and $F_w$ are two independent $1 \times 1$ convolutional transformations, $\sigma$ is the sigmoid function (which squashes outputs to the range $(0, 1)$ ), and the resulting tensors are $g^h \in \mathbb{R}^{C \times H \times 1}$ and $g^w \in \mathbb{R}^{C \times 1 \times W}$ .

What happens in detail: The split operation separates the $(H+W)$ -long spatial dimension of $f$ back into two segments: the first $H$ positions (which originated from $z^h$ ) become $f^h$ , and the remaining $W$ positions (which originated from $z^w$ ) become $f^w$ . The independent $1 \times 1$ convolutions $F_h$ and $F_w$ project from $C/r$ channels to $C$ channels—effectively, each learns to predict, from the compact encoded representation at a particular spatial position, how important each channel is at that spatial position.

Why split and use separate transformations: The two feature maps $f^h$ and $f^w$ encode fundamentally different types of information (vertical-position-aware features vs. horizontal-position-aware features). Using separate transformations $F_h$ and $F_w$ allows the network to learn different mappings for the two directions—for example, a channel might be important when active at the top of the image (high $h$ ) but less so when active at the bottom, and a different channel might be important when active on the left side. The alternative—using the same transformation for both—would force the horizontal and vertical attention to have identical channel-wise patterns, which would prevent the model from learning direction-specific attention weights.

The sigmoid activation at the output produces values in $(0, 1)$ , which are interpreted as continuous attention weights. A value near 1 means "attend strongly to this channel at this spatial position"; a value near 0 means "suppress this channel at this spatial position." The sigmoid is a standard choice for attention gating mechanisms because its output is bounded and smooth, allowing the attention weights to act as multiplicative gates.

Parameter count of the attention generation stage. The shared $1 \times 1$ convolution $F_1$ has $C \times (C/r) + (C/r)$ parameters (weights plus bias). Each of the two independent $1 \times 1$ convolutions $F_h$ and $F_w$ has $(C/r) \times C + C$ parameters. The total parameter count for the attention generation stage is approximately $2C^2/r + 2C/r + 2C$ , which is dominated by $2C^2/r$ . Compare this to SE attention's excitation stage, which has $2C^2/r$ parameters (one down-projection $T_1$ and one up-projection $T_2$ ). The coordinate attention block has approximately twice the bottleneck parameters due to the two independent up-projections ( $F_h$ and $F_w$ ), but the total is still linear in $C/r$ and typically adds less than 10% to the total backbone parameters (as the experimental results in Tables 2–5 demonstrate).

The reduction ratio $r$ controls the bottleneck size and therefore the parameter-accuracy tradeoff. A smaller $r$ means a larger bottleneck (more channels in the intermediate representation $f$ ), which preserves more information through the bottleneck but increases parameters in the $1 \times 1$ convolutions. The paper uses $r=32$ for default experiments and ablates $r=16$ (Table 4). With $r=16$ , the parameter count increases from 3.95M to 4.37M for MobileNetV2-1.0, while ImageNet top-1 accuracy improves from 74.3% to 74.7%, demonstrating that coordinate attention benefits from increased capacity in the bottleneck more than the competing methods (SE improves from 74.1% to 74.1% with the same $r$ reduction—zero additional gain, suggesting SE's channel-only encoding hits diminishing returns faster).

Coordinate-Aware Recalibration: Combining the Two Attention Maps

The final step applies the learned attention weights back onto the original input feature map through element-wise (Hadamard) multiplication:

$y_c(i, j) = x_c(i, j) \times g_c^h(i) \times g_c^w(j)$

where $y_c(i, j)$ is the output value at channel $c$ , row $i$ , column $j$ ; $x_c(i, j)$ is the original input value at the same position; $g_c^h(i)$ is the attention weight for channel $c$ at vertical position $i$ (a scalar); and $g_c^w(j)$ is the attention weight for channel $c$ at horizontal position $j$ (a scalar).

What it computes: For each channel $c$ and each spatial position $(i, j)$ , the output is the product of the original feature value and two attention weights—one encoding how strongly that row $i$ should be attended in channel $c$ , and one encoding how strongly that column $j$ should be attended in channel $c$ . The product $g_c^h(i) \times g_c^w(j)$ effectively creates a 2D attention map for each channel, but without ever explicitly computing or storing a full $H \times W$ matrix per channel—it is factorized into the outer product of two 1D vectors, which requires only $O(H + W)$ storage and computation per channel rather than $O(HW)$ .

Why this multiplicative form: The multiplication of row attention and column attention produces an attention map that is high only at positions where both the row and column are attended. If $g_c^h(i)$ is low (row $i$ is not important for channel $c$ ) and $g_c^w(j)$ is high (column $j$ is important), the product is low—the position $(i,j)$ is suppressed. Only when both are high does the position receive strong attention. This AND-like gating behavior means the network can learn to attend to rectangular regions (combinations of attended rows and attended columns) without needing to compute per-pixel attention values independently. The factorization assumes that attention is approximately separable into row-wise and column-wise components, which is an inductive bias that reduces parameters while still allowing the network to spatially localize objects.

The channel-wise nature of the attention: Note that the attention weights $g^h$ and $g^w$ are per-channel—different channels get different spatial attention patterns. Channel 17 might attend strongly to the top-left quadrant (high $g_{17}^h$ for small $i$ , high $g_{17}^w$ for small $j$ ), while channel 42 might attend to the center of the image. This is in contrast to CBAM's spatial attention, which produces a single spatial attention map shared across all channels (because it pools the channel dimension down to 1 before computing the spatial attention). The per-channel spatial attention is richer and allows the network to emphasize different spatial regions for different feature types.

Integration with the residual connection: In the typical integration pattern (Figure 3), the coordinate attention block output replaces the original feature map within a residual block. The overall output of the block is either the attention output directly (if the attention block is placed inside the residual branch) or the attention output added to a skip connection (if placed before the residual addition). The paper does not introduce a separate residual connection for the attention block itself—it integrates into the existing residual structure of the host architecture, following the same pattern as SE and CBAM integration.

Integration into Mobile Network Architectures

The paper demonstrates coordinate attention integration into two fundamentally different mobile network building blocks: the inverted residual block from MobileNetV2 (Sandler et al., 2018) and the sandglass bottleneck block from MobileNeXt (Zhou et al., 2020). These represent two distinct architectural paradigms for mobile networks, and the choice to demonstrate integration into both reinforces the claim that coordinate attention is a general-purpose attention mechanism, not one coupled to a specific backbone design.

Integration into the inverted residual block (Figure 3a). The standard inverted residual block has three layers: a $1 \times 1$ pointwise convolution that expands the channel dimension (from $C_{\text{in}}$ to $t \times C_{\text{in}}$ where $t$ is the expansion factor, typically 6), a $3 \times 3$ depthwise convolution that operates on each channel independently with spatial filtering, and a $1 \times 1$ pointwise convolution that projects back to $C_{\text{out}}$ . A residual connection (skip connection) connects the input to the output when $C_{\text{in}} = C_{\text{out}}$ and the stride is 1.

The coordinate attention block is inserted after the depthwise convolution and before the final $1 \times 1$ projection, at the position where the feature maps have the expanded channel dimension $t \times C_{\text{in}}$ and have already undergone spatial filtering. This is the same insertion point used for SE attention and CBAM in the MobileNetV2 + SE / + CBAM experiments. The rationale for this position: the depthwise convolution has already processed spatial information locally (within a $3 \times 3$ neighborhood), and the attention mechanism can now recalibrate these spatially-filtered features based on global context before the channel projection compresses them. The attention block operates on the expanded-dimension features, which gives it a richer channel space to learn attention patterns over.

Integration into the sandglass bottleneck block (Figure 3b). The sandglass block has a reversed structure compared to the inverted residual: two $3 \times 3$ depthwise convolutions at the ends (operating on lower-dimensional features) and two $1 \times 1$ pointwise convolutions in the middle that project to a higher-dimensional latent space and back. The coordinate attention block is inserted after the first depthwise convolution, before the first $1 \times 1$ pointwise convolution, again operating on features after initial spatial filtering but before channel transformation. An additional attention block is placed after the second $1 \times 1$ pointwise convolution and before the final depthwise convolution, performing attention in the compressed channel space before spatial filtering.

This dual-insertion pattern in the sandglass block demonstrates that coordinate attention can be applied multiple times within a single block at different channel dimensionalities—once in the input's channel space and once in the bottleneck's channel space—without bloating the parameter count.

The general principle for insertion position: In both architectures, attention blocks are placed at positions where (1) the feature maps carry meaningful spatial structure (not after global average pooling or at the fully-connected layer stage) and (2) the channel dimension is sufficiently large for the attention mechanism to learn meaningful channel interdependencies (not at extreme bottlenecks where $C$ is very small). The paper does not insert attention blocks at every possible position—following the practice established by SE attention, they are placed at strategic locations (typically every block or every other block, depending on the architecture and weight multiplier) to balance cost and benefit.

Summary of Design Choices and Their Justifications

1D pooling factorization over 2D pooling: Preserves positional information along one spatial dimension per branch while capturing long-range dependencies along the orthogonal dimension. A 2D global pooling would collapse both dimensions simultaneously (SE's limitation); a full non-local operation would capture all pairs but at quadratic cost. The 1D factorization hits a computational sweet spot: $O(H+W)$ cost per channel rather than $O(HW)$ .

Shared $1 \times 1$ convolution followed by independent transformations: The shared convolution learns a compact common encoding of spatial-direction information, enabling information sharing between horizontal and vertical branches. The independent transformations then allow direction-specific attention weight prediction—horizontal importance patterns need not mirror vertical patterns for the same channel. The alternative of fully shared or fully independent transformations throughout would lose either the parameter efficiency of sharing or the expressiveness of direction-specificity.

Multiplicative combination of row and column attention weights: Produces a factorized 2D attention map that can attend to rectangular regions without storing an $H \times W$ matrix. The outer product $g^h(i) \times g^w(j)$ is an AND-like gate—attention is high only where both row and column are attended—which provides spatially selective recalibration while assuming approximate separability of spatial attention. This inductive bias is appropriate for many visual objects that have approximately axis-aligned spatial extent (faces, vehicles, buildings), though it would struggle with attention patterns that are fundamentally non-separable (e.g., diagonal structures).

Per-channel spatial attention over channel-shared spatial attention: Unlike CBAM which produces a single spatial attention map for all channels (by reducing the channel dimension to 1–2 before spatial processing), coordinate attention produces different spatial attention patterns for different channels. This is important because different channels detect different features (edges at various orientations, textures, colors, object parts), and these features are not spatially co-located—an edge detector for vertical edges should attend to different image regions than a texture detector for fur patterns. Channel-shared spatial attention forces all feature types to share the same "where to look" signal, losing this specificity.

Integration at existing attention insertion points: Coordinate attention is designed as a drop-in replacement for SE attention and CBAM, not as a new architecture requiring custom block design. This design choice maximizes practical utility—any existing mobile network that uses SE blocks can swap in coordinate attention blocks with minimal code changes, simply replacing the attention module while keeping the backbone architecture intact. The experimental section validates this by showing consistent gains when replacing SE with coordinate attention in MobileNetV2, MobileNeXt, and EfficientNet without any other architectural modifications.

4. Key Insights and Innovations

Innovation 1: Spatial Attention Through Dimensional Factorization as a Third Path Between Local Convolutions and Quadratic Self-Attention

The paper's most conceptually distinctive move is identifying and occupying a previously unnamed point in the design space of spatial attention mechanisms. Before this work, the field implicitly operated under a dichotomy: spatial attention was either local (captured by convolutions with fixed kernel sizes, as in CBAM's $7 \times 7$ convolution) or global-and-quadratic (captured by non-local/self-attention blocks that compute pairwise relationships between all spatial positions, costing $O(H^2 W^2)$ ). The local approach is cheap but myopic—a single convolution layer cannot relate distant pixels. The global approach captures true long-range dependencies but is computationally prohibitive for mobile deployment, as the paper notes in Section 2.2: self-attention modules "are often adopted in large models but not suitable for mobile networks."

Coordinate attention introduces a third category: global spatial encoding along one dimension combined with precise positional preservation along the orthogonal dimension, achieved by factorizing a single 2D operation into two complementary 1D operations. This is not a compromise that averages the two extremes—it is a structurally different way to encode spatial information that achieves properties neither extreme provides alone. A horizontal 1D global pooling (Eqn. 4) has a receptive field spanning the entire width of the image—all columns contribute to the value at each row—making it genuinely "long-range" along the horizontal axis. Simultaneously, it preserves exact vertical position because pooling operates only across columns, leaving the row index intact. The vertical pooling provides the symmetric property.

This factorization reframes what "long-range spatial attention" means for mobile networks. Rather than asking "how can we approximate full pairwise attention cheaply?" (the approach of low-rank or sparse attention variants), coordinate attention asks "which long-range spatial relationships matter most, and can we capture them by decomposing along axes?" The answer—that row-wise and column-wise global context, combined multiplicatively, is sufficient for substantial gains on dense prediction tasks—is an empirical finding that challenges the assumption that per-pixel pairwise attention is necessary for spatial reasoning. The Cityscapes segmentation results (Table 9: +2.6 mIoU over SE attention, matching or exceeding what might be expected from far more expensive attention mechanisms) provide the cleanest evidence that axis-decomposed global context captures meaningful spatial structure.

This is a fundamental conceptual contribution rather than an incremental refinement because it opens a new axis (no pun intended) for designing efficient attention: dimensional factorization. Future work can explore whether other factorizations (e.g., radial+angular in polar coordinates, or multi-scale 1D strips at different resolutions) provide similar benefits, building on the core insight that global context need not be isotropic to be useful.

Innovation 2: Diagnosing and Exploiting the Classification-Dense Prediction Gap as an Attention Design Principle

The paper makes a diagnostic move that is as important as its architectural contribution: it identifies that the gap between an attention mechanism's performance on ImageNet classification and its performance on dense prediction tasks (detection, segmentation) reveals what kind of information the attention mechanism encodes. This is not merely an observation that "some methods transfer better than others"—it is a principled argument that classification-to-segmentation transfer gap serves as a litmus test for spatial encoding quality.

The evidence for this diagnostic claim is structured across the paper's three evaluation settings. On ImageNet classification (Table 2, MobileNetV2-1.0): SE achieves 73.5% (+1.2 over baseline), CBAM achieves 73.6% (+1.3), and coordinate attention achieves 74.3% (+2.0). The differences are modest at the scale of classification accuracy—all three attention methods provide a boost, and the ordering is SE ≈ CBAM < CA. On COCO object detection (Table 6): SE achieves 23.7 AP (+1.4 over baseline), CBAM achieves 23.0 (+0.7—actually worse than SE), and coordinate attention achieves 24.5 (+2.2). The gap between coordinate attention and the alternatives widens in absolute terms and CBAM's advantage over SE disappears. On Pascal VOC semantic segmentation (Table 8, output stride 16): SE achieves 71.69% mIoU (+0.85 over baseline), CBAM achieves 71.28% (+0.44), and coordinate attention achieves 73.32% (+2.48). The relative advantage of coordinate attention over SE grows from roughly 1.1× on classification (74.3/73.5) to roughly 1.03× on detection and 1.02× on segmentation—but more importantly, the absolute gap between coordinate attention and SE increases from 0.8 percentage points on classification to 1.6 points on segmentation at output stride 16, and to 1.4 points at output stride 8 (73.96 vs. 72.52).

The paper interprets this pattern explicitly in Section 4.4.2:

"We argue that this is because our coordinate attention is able to capture long-range dependencies with precise positional information, which is more beneficial to vision tasks with dense predictions, such as semantic segmentation."

This interpretation transforms the transfer gap from a mere performance metric into a design validation signal. If an attention mechanism claims to encode positional information, its benefits should be most pronounced on tasks that require spatial precision. The fact that SE attention—which the paper mathematically demonstrates discards all positional information via 2D global pooling—shows diminishing relative returns as spatial precision demands increase, while coordinate attention shows growing relative returns, provides a coherent explanation grounded in the mechanisms' mathematical properties.

This is a conceptual contribution that extends beyond the specific method. The paper effectively argues that the research community should evaluate lightweight attention mechanisms not just on ImageNet top-1 (where channel-only attention already does well) but on a spectrum of tasks with varying spatial precision requirements, because the shape of the performance curve across tasks reveals what the attention mechanism actually learns to encode. This is a methodological insight—a diagnostic framework—that future attention mechanism papers can adopt, regardless of whether they use coordinate attention's specific factorization approach.

Innovation 3: The Sufficiency of Axis-Aligned Spatial Factorization for Object Localization

The paper makes an implicit but empirically validated claim about visual structure: that row-wise and column-wise global context, when combined multiplicatively, provides sufficient spatial information to substantially improve object localization in mobile networks. This is not obvious a priori. One could reasonably argue that factorizing spatial attention into separable horizontal and vertical components would fail on objects with complex non-axis-aligned shapes—diagonal structures, curved boundaries, or occluded objects where the "important rows" and "important columns" do not form clean rectangular regions.

The experimental results, particularly on segmentation tasks where precise boundary localization matters, suggest that this concern is outweighed in practice by the benefits of global context along each axis. The multiplicative gating $g^h(i) \times g^w(j)$ creates an attention map that can suppress background regions even when they share rows or columns with foreground objects: a background pixel at position $(i,j)$ will only be attended if both the row $i$ and column $j$ have high attention weights, which is less likely for background positions that typically don't share both coordinates with foreground objects. This is a form of implicit spatial grouping achieved without explicit pairwise computation.

This finding is significant because it challenges an unspoken assumption in the attention literature: that modeling spatial relationships requires computing interactions between spatial positions (either through convolutions that relate neighboring positions or through attention mechanisms that relate all pairs). Coordinate attention demonstrates that aggregating along orthogonal axes and recombining multiplicatively is a viable alternative—one that captures enough spatial structure to matter for practical tasks while remaining computationally tractable for mobile devices. The Cityscapes result in Table 9 (74.0% mIoU with coordinate attention vs. 71.4% for vanilla MobileNetV2, with only 0.5M additional parameters) makes this case concretely: the network can segment urban street scenes—which contain objects at various scales, orientations, and positions—substantially better with only axis-aligned global context added.

This is an incremental contribution in the sense that it builds on the well-established idea of separable convolutions (factorizing a $k \times k$ convolution into $k \times 1$ and $1 \times k$ ), but it is a conceptual leap in that separable convolutions are local operations while coordinate attention's 1D pooling operations are global. The paper extends the factorization idea from local filtering to global context aggregation, which is a non-trivial generalization. The result that this works well for attention—not just for convolution—expands the design vocabulary for efficient spatial mechanisms.

Innovation 4: Reinterpreting Channel Attention Bottlenecks as Position-Aware Feature Encoders

A subtle but important conceptual move in the paper is how it reuses the channel attention bottleneck structure (the reduce-and-expand pattern from SE attention's excitation stage) for a fundamentally different purpose. In SE attention, the bottleneck network $T_1$ and $T_2$ (Eqn. 3) learns a mapping from global channel statistics to channel importance weights—it answers the question "given how active each channel is globally, which channels should be emphasized?" This is a pure channel-wise computation: the spatial information has already been destroyed by the 2D pooling before the bottleneck sees the data.

In coordinate attention, the shared $1 \times 1$ convolution $F_1$ (Eqn. 6) operates on the concatenated direction-aware feature maps $[z^h, z^w]$ , which contain $(H+W)$ spatial positions each encoded with $C$ channels of information. The bottleneck $F_1$ reduces this to $C/r$ channels but preserves the $(H+W)$ spatial positions—it is a $1 \times 1$ convolution that applies the same transformation independently at each spatial position, not a fully-connected layer that mixes positions. This means the bottleneck is learning a position-invariant encoding of direction-aware features: "given a channel's activation pattern along a row (or column), what compact representation best captures its spatial significance?"

The conceptual innovation is that the bottleneck has been repurposed from a channel-relationship modeler (SE) to a position-aware feature encoder. The subsequent independent transformations $F_h$ and $F_w$ then decode this compact representation into channel-specific attention weights that vary spatially. The bottleneck thus serves as an information bottleneck in the proper sense—it forces the network to learn a compressed representation of spatial-direction features that must be sufficient to predict both horizontal and vertical attention patterns.

This reinterpretation matters because it shows that the bottleneck structure, which was originally introduced in SE attention purely for parameter efficiency (reducing $2C^2$ parameters to $2C^2/r$ ), can serve a more fundamental representational role when the input to the bottleneck carries spatial structure. The paper does not explicitly frame this as an innovation—it presents it as a straightforward design choice—but recognizing that the same architectural pattern serves qualitatively different functions depending on what information survives to reach it is a conceptual contribution that informs future attention block design. The ablation in Table 1, showing that either horizontal-only or vertical-only attention matches SE performance while combined coordinate attention substantially exceeds it, provides evidence that the bottleneck is learning to jointly encode both directional signals rather than simply memorizing channel importance.

This is an incremental refinement of the SE bottleneck design, but the shift in what the bottleneck operates on—from a spatially-collapsed channel vector to spatially-structured direction-aware feature maps—fundamentally changes its role in the network, making it a more significant conceptual departure than the architectural similarity might suggest.

5. Experimental Analysis

Evaluation Methodology

Dataset. ImageNet ILSVRC 2012 for classification (1.28M training images, 50K validation images, 1000 classes). MS COCO 2017 for object detection (118K training, 5K validation). Pascal VOC 2012 for both detection (VOC 2007 test set with 4,952 images) and semantic segmentation (1,464 training, 1,449 validation, augmented with extra annotations from Hariharan et al. to 10,582 training images). Cityscapes for semantic segmentation (2,975 fine-annotated training images, 500 validation images, 19 classes). These benchmarks span the spectrum from single-label classification to dense per-pixel prediction, which is deliberate: the paper's central claim is that coordinate attention's spatial encoding benefits scale with the spatial precision demands of the task.
Base model(s). Three mobile network families spanning different architectural paradigms: MobileNetV2 (Sandler et al., 2018) with inverted residual blocks, MobileNeXt (Zhou et al., 2020) with sandglass bottleneck blocks, and EfficientNet-b0 (Tan and Le, 2019) which is NAS-derived and includes SE attention by default. The weight multipliers {1.0, 0.75, 0.5} are evaluated for the first two families to test across model scales—from roughly 3.5M parameters down to 2.0M. The choice of three architecturally distinct families is a deliberate robustness check: if coordinate attention only worked well with inverted residuals, its claimed generality would be suspect.
Metrics. ImageNet: top-1 accuracy (%) on the 50K validation set using single-crop testing. COCO detection: AP (averaged over IoU thresholds 0.50:0.95), AP50, AP75, and AP across small/medium/large object scales (APS, APM, APL). Pascal VOC detection: mAP (%). Semantic segmentation: mean Intersection-over-Union (mIoU, %) on both Pascal VOC 2012 and Cityscapes validation sets. All metrics are standard for their respective tasks, enabling direct comparison with prior work.
Baselines. Three categories: (1) No-attention baselines — vanilla MobileNetV2, MobileNeXt, and EfficientNet-b0 without any attention modules beyond what the architecture natively includes (EfficientNet-b0 has SE attention built in, so the baseline for that family is the original SE-equipped model). (2) Channel-only attention — the SE attention block (Hu et al., 2018), which serves as the primary reference point since it is "the most popular attention mechanism for mobile networks" (Section 1). (3) Channel+spatial attention — CBAM (Woo et al., 2018), which adds a spatial attention module after channel attention using a 7×7 convolution on channel-pooled features. For EfficientNet-b0 comparisons (Table 5), additional NAS-derived baselines are included: PNAS, DARTS, ProxylessNAS, AmoebaNet-A, FBNet-C, MnasNet-A3. For detection (Table 6), MobileNetV1, MobileNetV3, and MnasNet-A1 serve as architecture-reference baselines.
Generation budget / compute accounting. The paper measures computational cost using two metrics: number of parameters (in millions, M) and multiply-add operations (M-Adds, in millions or billions depending on context). For mobile deployment, on-device latency (in milliseconds, measured on a Google Pixel 4 device) is additionally reported in the main ablation (Table 1). All methods are compared at approximately equal parameter counts and M-Adds—the attention modules are inserted at identical positions in the backbone architectures, and any parameter increase from the attention block is explicitly quantified. This is a fair comparison protocol because it prevents a method from winning simply by using more computation; the efficiency constraint is binding.
Cross-validation / statistical protocol. No cross-validation or statistical significance testing is reported explicitly. The standard ImageNet protocol is followed: training on the 1.28M training set, validation on the 50K validation set. For semantic segmentation, the standard train/val splits of Pascal VOC 2012 and Cityscapes are used. All results are single-model, single-crop, with no test-time augmentation or model ensembling (stated explicitly in Tables 6, 8, and 9). The absence of error bars or multiple random seeds is a limitation—particularly for the smaller Pascal VOC and Cityscapes validation sets (1,449 and 500 images respectively), where variability in mIoU could be non-trivial. However, the consistency of the pattern across three tasks, multiple datasets, and multiple backbone architectures provides informal robustness.

Main Quantitative Results

ImageNet Classification: Modest but Consistent Gains Across Architectures and Scales

The headline classification result is in Table 2 for MobileNetV2-1.0: coordinate attention achieves 74.3% top-1 accuracy, compared to 73.5% for SE attention (+0.8 percentage points), 73.6% for CBAM (+0.7 points over coordinate attention's baseline comparison, though within 0.1 of SE), and 72.3% for the no-attention baseline. This +0.8 point gain over SE attention is the central ImageNet claim: coordinate attention modestly but consistently outperforms the dominant mobile attention mechanism on the standard classification benchmark.

The pattern holds across all three weight multipliers tested for MobileNetV2 (Table 2): at 0.75× width, coordinate attention achieves 72.1% vs. 71.5% for both SE and CBAM (+0.6); at 0.5× width, 67.0% vs. 66.4% for both SE and CBAM (+0.6). The absolute gain is roughly constant (0.6–0.8 points) across a 2× range of model sizes (3.5M to 2.0M parameters), suggesting the benefit does not diminish as the model shrinks—an important property for mobile deployment where smaller variants are often used.

For MobileNeXt (Table 3), the pattern replicates: at 1.0×, coordinate attention achieves 75.2% vs. 74.7% for SE (+0.5) and 74.0% for no-attention baseline; at 0.75×, 73.2% vs. 72.6% (+0.6); at 0.5×, 69.4% vs. 68.7% (+0.7). The gain over SE is slightly smaller than on MobileNetV2 (0.5–0.7 vs. 0.6–0.8), but the consistent advantage across two fundamentally different block structures (inverted residual vs. sandglass) demonstrates that the attention mechanism's benefit is not tied to a specific residual pattern.

The EfficientNet-b0 result (Table 5) tests whether coordinate attention can improve a stronger, NAS-derived baseline that already includes SE attention. Coordinate attention achieves 76.9% top-1, compared to 76.3% for the original EfficientNet-b0 with SE—a +0.6 point gain. This is particularly notable because EfficientNet-b0 was already highly optimized through architecture search, and the SE attention positions were part of the searched architecture. Replacing SE with coordinate attention without modifying any other aspect of the searched design yields a nontrivial improvement, suggesting the gain comes from the attention mechanism itself rather than from interactions with manually-designed block structures.

A comparison across all architectures and weight multipliers reveals a consistent ordering: no attention < SE attention ≈ CBAM < coordinate attention. CBAM and SE are statistically indistinguishable on classification (73.5 vs. 73.6 for MobileNetV2-1.0; 71.5 vs. 71.5 for MobileNetV2-0.75; 66.4 vs. 66.4 for MobileNetV2-0.5), which the paper interprets as evidence that CBAM's spatial attention module does not contribute meaningfully for mobile networks on classification. The paper states this explicitly in Section 4.3: "it seems that its spatial attention module shown in Figure 2(b) does not contribute in mobile networks compared to the SE attention."

The latency measurements in Table 1 provide the on-device cost characterization: the baseline MobileNetV2 runs at 14–16ms on a Google Pixel 4; adding SE attention increases this to 16–18ms; adding coordinate attention increases it to 17–19ms. The additional 1–3ms over SE attention represents the cost of the two 1D pooling operations, the concatenation, the shared 1×1 convolution, the split, and the two independent 1×1 convolutions. The paper presents this as "nearly no computational overhead" (Abstract), though a 1–3ms increase on a 16ms baseline represents a 6–19% relative increase in latency—noticeable but arguably acceptable given the accuracy gains, particularly on downstream tasks.

Object Detection on COCO: The Gap Over SE Attention Widens

The COCO detection results in Table 6 use the SSDLite320 detector with MobileNetV2 backbones. The key numbers: coordinate attention achieves 24.5 AP, compared to 22.3 AP for the no-attention baseline (+2.2), 23.7 AP for SE attention (+0.8 over SE), and 23.0 AP for CBAM (+1.5 over CBAM). The AP gain over the baseline is larger than on ImageNet (+2.2 vs. +2.0 absolute points), but more importantly, the gain over SE attention specifically has grown: from +0.8 on ImageNet to +0.8 on COCO AP—which is actually the same absolute gain, but the relative improvement over SE is smaller because SE itself has gained less on detection relative to the baseline.

Breaking down by object scale (Table 6, columns APS, APM, APL): coordinate attention achieves 2.3 APS, 26.2 APM, and 45.9 APL, compared to SE attention at 2.2 APS, 25.4 APM, and 44.7 APL. The gains are distributed across all object scales, with the largest absolute improvement at medium objects (+0.8 APM). This is consistent with the interpretation that coordinate attention helps localize objects—medium objects are large enough to span multiple rows and columns (making row/column attention meaningful) but small enough that precise localization remains challenging, whereas large objects are already easy to detect and small objects may be too small for the 1D attention maps to provide discriminative spatial information.

The comparison with NAS-derived models in Table 6 is instructive: MobileNetV3 (NAS-derived) achieves 22.0 AP with 5.0M parameters and 0.62B M-Adds; MnasNet-A1 (NAS-derived) achieves 23.0 AP with 4.9M parameters and 0.8B M-Adds; MobileNetV2 + coordinate attention achieves 24.5 AP with 4.8M parameters and 0.8B M-Adds. The hand-designed network with coordinate attention outperforms both NAS-derived architectures at comparable or lower computational cost, which is a strong practical result—it suggests that a well-designed attention mechanism can substitute for expensive architecture search, at least for detection transfer.

CBAM's detection performance (23.0 AP) is notably below SE's (23.7 AP), and actually matches the MnasNet-A1 baseline. This is a striking negative result for CBAM: its spatial attention module, which uses a 7×7 convolution on channel-pooled features, not only fails to help but actually hurts detection performance relative to SE attention (which has no spatial attention at all). The paper does not deeply analyze this failure, but it is consistent with the interpretation that channel squeezing (compressing C channels to 2 before spatial processing) loses too much information for the spatial attention to be beneficial, and the local 7×7 convolution cannot capture the long-range spatial context needed for detection.

Object Detection on Pascal VOC: Coordinate Attention Alone Improves, SE and CBAM Do Not

The Pascal VOC 2007 detection results in Table 7 reveal a pattern even starker than COCO: coordinate attention achieves 73.1% mAP, compared to 71.7% for vanilla MobileNetV2 (+1.4), while both SE attention and CBAM achieve exactly 71.7% mAP (+0.0)—they provide zero improvement over the no-attention baseline on this benchmark with this detector. This is the paper's most dramatic result for the detection task and supports the central claim that spatial information matters more for some tasks than others: Pascal VOC has fewer classes (20) with more distinctive spatial layouts and larger typical object sizes than COCO, making spatial attention potentially more impactful when it works, but also revealing that channel-only attention (SE) or channel-squeezed spatial attention (CBAM) contributes nothing.

The fact that SE attention, which provides +1.4 AP on COCO and +1.2% top-1 on ImageNet, provides zero mAP gain on Pascal VOC detection is not explained in the paper but is a significant observation. It suggests that the channel recalibration provided by SE attention may be redundant or already implicitly learned when the backbone is fine-tuned on a detection task with sufficient data (COCO has 118K training images; Pascal VOC has considerably fewer, so data quantity alone cannot explain the difference). The detector architecture (SSDLite320) is the same across both benchmarks, ruling out detector-specific interactions. This inconsistency—SE helping on COCO but not Pascal VOC—is not discussed by the authors and represents an unexplained empirical result that merits investigation.

Semantic Segmentation on Pascal VOC 2012: The Largest Gains, Consistent with Spatial Information Hypothesis

The semantic segmentation results in Table 8, using DeepLabV3 with MobileNetV2 backbones, show the largest absolute gains for coordinate attention. At output stride 16: coordinate attention achieves 73.32% mIoU, compared to 70.84% for vanilla MobileNetV2 (+2.48), 71.69% for SE attention (+1.63 over SE), and 71.28% for CBAM (+2.04 over CBAM). At output stride 8 (higher resolution, more spatial detail preserved): coordinate attention achieves 73.96% mIoU vs. 72.52% for SE (+1.44) and 71.82% for vanilla (+2.14).

The gain over SE attention is +1.63 mIoU at stride 16 and +1.44 mIoU at stride 8—substantially larger than the +0.8 point gain on ImageNet classification. This is the pattern the paper emphasizes in its closing discussion (Section 5): the benefits of coordinate attention scale with the spatial precision demands of the task. Classification asks "what object is in this image?" and can succeed with spatially-invariant features. Segmentation asks "which pixels belong to this object?" and requires precise spatial localization. An attention mechanism that encodes positional information should show larger gains on the latter, and coordinate attention does (2.48 mIoU over baseline vs. 2.0 top-1 over baseline on ImageNet; 1.63 mIoU over SE on segmentation vs. 0.8 top-1 over SE on classification).

CBAM's segmentation performance (71.28% at stride 16) is again below SE's (71.69%), and actually approaches the no-attention baseline (70.84%). This reinforces the detection finding: CBAM's spatial attention module, far from helping, degrades performance when fine-tuned on dense prediction tasks. The paper attributes this to two factors (Section 4.3): channel squeezing to 2 channels losing information, and the 7×7 convolution only capturing local context. The segmentation results provide the strongest evidence that these limitations are not merely theoretical but cause measurable performance degradation on tasks requiring global spatial reasoning.

At output stride 8, all methods improve as expected (higher resolution preserves more spatial detail for the segmentation head), and the relative ordering remains consistent: coordinate attention > SE attention > CBAM > no attention. The coordinate attention gain over SE shrinks slightly at the higher resolution (1.44 vs. 1.63 mIoU), which might indicate that the attention mechanism's global spatial encoding is somewhat more valuable when the feature maps are coarser (stride 16) and long-range context is harder to capture through the backbone convolutions alone. At stride 8, local convolutions in the backbone and ASPP module may partially compensate for the lack of explicit spatial attention, reducing coordinate attention's relative advantage.

Semantic Segmentation on Cityscapes: Replication on a Different Domain

The Cityscapes results in Table 9 replicate the segmentation findings on a different domain (urban street scenes vs. Pascal VOC's mix of natural images). At output stride 8 with full-resolution (1024×2048) testing: coordinate attention achieves 74.0% mIoU, compared to 71.4% for vanilla MobileNetV2 (+2.6), 72.2% for SE attention (+1.8 over SE), and 71.4% for CBAM (+2.6 over CBAM, which again fails to improve over the no-attention baseline). The gain over SE attention (+1.8 mIoU) is consistent with the Pascal VOC segmentation results (+1.63 at stride 16, +1.44 at stride 8), providing cross-dataset validation of the claim that coordinate attention's spatial encoding benefits dense prediction.

CBAM's complete failure on Cityscapes (71.4% mIoU, identical to the no-attention baseline) is the most damning evidence against its spatial attention module for mobile networks. On a dataset where spatial context is critical—urban scenes with cars, pedestrians, traffic signs, and buildings at various scales and positions—CBAM's local 7×7 spatial attention provides no benefit whatsoever, while coordinate attention's global 1D encoding yields a +2.6 mIoU improvement. This asymmetry is exactly what the paper's mechanistic analysis predicts: local spatial context (7×7 convolution) is insufficient for segmentation tasks requiring long-range spatial reasoning; global context along orthogonal axes (1D pooling) captures enough spatial structure to matter.

The +2.6 mIoU gain on Cityscapes, with only 0.5M additional parameters over the vanilla MobileNetV2, is arguably the paper's strongest single result for practical impact. Cityscapes is the standard benchmark for real-world mobile segmentation applications (autonomous driving, mobile mapping), and a 2.6 mIoU improvement represents substantial progress on a task where gains are typically hard-won and expensive in compute.

Ablation Studies and Robustness Checks

Importance of encoding both spatial directions (horizontal + vertical vs. single direction): Table 1 ablates the core design choice. Adding only horizontal attention (X Attention) achieves 73.5% top-1 on ImageNet—identical to SE attention's 73.5%. Adding only vertical attention (Y Attention) also achieves 73.5%. Combining both into full coordinate attention achieves 74.3% (+0.8 over single-direction variants or SE). This ablation demonstrates that: (1) a single 1D pooling direction provides no advantage over 2D global pooling (SE) on classification, suggesting the benefit of 1D pooling comes from having both directions working complementarily; (2) the improvement from coordinate attention over SE is entirely due to the conjunction of horizontal and vertical encodings, not from any single-direction advantage. The paper does not ablate whether two independent 1D attention branches without the shared $1 \times 1$ convolution would perform similarly, which would test whether the concatenation-and-split design matters.

Impact of the reduction ratio r: Table 4 compares r=32 (default, 3.95M parameters) with r=16 (4.37M parameters). Coordinate attention at r=32 achieves 74.3% top-1; at r=16 improves to 74.7% (+0.4). SE attention at r=24 achieves 73.5%; at r=12 achieves 74.1% (+0.6). CBAM at r=24 achieves 73.6%; at r=12 achieves 74.1% (+0.5). The key observation is that coordinate attention at r=32 (74.3%) already outperforms SE at r=12 (74.1%) and CBAM at r=12 (74.1%), despite having fewer parameters (3.95M vs. 4.28M). When coordinate attention is given comparable parameters (r=16, 4.37M), the gap widens to +0.6 over SE and CBAM at their best reduction ratios. This demonstrates that coordinate attention's advantage is not simply due to having more parameters—it is more parameter-efficient, achieving better accuracy with fewer parameters.

Robustness across weight multipliers and architectures: Table 2 and Table 3 collectively ablate the interaction between coordinate attention and model scale. The consistent +0.5 to +0.8 point gain over SE attention across weight multipliers {0.5, 0.75, 1.0} and across two architectures (MobileNetV2 and MobileNeXt) demonstrates that the benefit does not depend on a specific model capacity or block structure. The gain is remarkably stable: 0.6–0.8 for MobileNetV2, 0.5–0.7 for MobileNeXt. There is no systematic trend where coordinate attention's relative advantage grows or shrinks with model size—it is roughly constant.

Grad-CAM visualization (Figure 4): While not a quantitative ablation, Figure 4 provides qualitative evidence for the paper's claim that coordinate attention "can more precisely locate the objects of interest." The Grad-CAM visualizations show feature maps before and after each attention block for five images. For the SE attention row, the post-SE feature maps show modest refinement of the pre-SE maps—the attended regions are slightly more focused but the overall activation pattern is similar. For CBAM, the post-CBAM maps show some spatial concentration but also apparent noise or mislocalization. For coordinate attention, the post-CA maps consistently show tighter localization around the target object (leaf beetle, flamingo, screen, beer glass, black bear) with less activation on background regions. This qualitative evidence supports the quantitative finding that coordinate attention helps spatial localization, though Grad-CAM has known limitations and should be interpreted cautiously.

Transfer learning robustness across tasks and datasets: The paper implicitly ablates transfer learning robustness by evaluating the same ImageNet-pretrained backbones on three different downstream tasks (ImageNet classification, COCO detection, Pascal VOC detection, Pascal VOC segmentation, Cityscapes segmentation). The consistent ordering—coordinate attention > SE attention > CBAM for dense tasks—across five different evaluation settings (ImageNet val, COCO val, VOC test, VOC val, Cityscapes val) is a strong robustness check. If the coordinate attention benefit were specific to ImageNet's class distribution or to classification objectives, it would not transfer to detection and segmentation.

Critical Assessment

Does coordinate attention genuinely encode "precise positional information" as claimed? The experiments provide indirect evidence through improved performance on spatially-demanding tasks (segmentation mIoU, detection AP) and Grad-CAM visualizations. However, the paper does not directly measure whether the learned attention maps $g^h$ and $g^w$ actually correspond to object positions, nor does it probe whether the multiplicative gating $g^h(i) \times g^w(j)$ produces attention maps that align with object bounding boxes or segmentation masks. A direct evaluation—for example, measuring the overlap between the product attention map and ground-truth object masks on a segmentation dataset—would provide stronger evidence for the "precise positional information" claim. The current evidence shows that coordinate attention helps tasks requiring spatial precision, which is consistent with the claim but does not prove the specific mechanism.

The claim that coordinate attention outperforms SE and CBAM is strongly supported but with a nuance about CBAM. In every experiment across all three tasks, coordinate attention achieves higher accuracy than both SE and CBAM. However, the CBAM baseline is consistently the weakest of the three attention mechanisms—matching or underperforming SE on most benchmarks, and completely failing (zero improvement over no-attention) on VOC detection and Cityscapes segmentation. This means coordinate attention's advantage over CBAM, while real, is over a relatively weak competitor. The more meaningful comparison is against SE attention (the dominant mobile attention mechanism), where coordinate attention shows consistent but modest gains on classification (+0.5 to +0.8 top-1) and more substantial gains on dense prediction (+1.4 to +1.8 mIoU on segmentation). The paper's framing around outperforming CBAM may overstate the practical significance—practitioners would more likely be choosing between coordinate attention and SE attention, where the tradeoff is accuracy vs. implementation simplicity.

A genuine weakness: the single-model, single-run evaluation protocol. All ImageNet results are from a single training run per configuration; all detection and segmentation results are from fine-tuning a single pretrained checkpoint. There are no error bars, no multiple random seeds, and no cross-validation. For ImageNet with its 50K validation images, the variability in top-1 accuracy across runs with different random seeds is typically small (0.1–0.2%) but not zero, and some of the reported differences (e.g., coordinate attention at 74.3% vs. SE at 73.5%—a 0.8 point gap) could be partially influenced by training noise. For Pascal VOC segmentation with only 1,449 validation images, mIoU variability across fine-tuning runs could be larger (0.5–1.0 mIoU is not uncommon in segmentation literature). Reporting mean and standard deviation over multiple runs, or at minimum noting that results are from a single run, would strengthen the reliability of the claims.

The absence of important baselines and ablation targets. The paper does not compare against several attention mechanisms that would help locate coordinate attention in the design space: (1) a simple baseline that adds $1 \times H$ and $W \times 1$ average pooling directly as spatial feature concatenation without the bottleneck network, to test whether the bottleneck and attention gating mechanism matters or whether the raw pooled features are sufficient; (2) a baseline using only global max pooling instead of average pooling for the 1D encodings; (3) any variant of the non-local/self-attention mechanisms that the paper discusses as motivation but never empirically benchmarks (e.g., a lightweight axial attention variant that factorizes self-attention along rows and columns, which would be the closest conceptual competitor); (4) the impact of inserting coordinate attention at different positions within the residual blocks, as opposed to the single insertion point used throughout. These missing experiments would clarify which aspects of the design contribute to performance. In particular, the absence of an axial attention baseline is notable because axial attention also factorizes spatial attention along height and width dimensions, and directly comparing coordinate attention against axial attention would test whether the 1D pooling approach is genuinely superior to factorized self-attention for mobile networks, or whether both capture similar spatial structure.

The EfficientNet-b0 result is promising but limited. The paper replaces SE with coordinate attention in EfficientNet-b0 and sees a +0.6 top-1 improvement (Table 5). However, EfficientNet's architecture was searched with SE attention in place—the positions of SE blocks, the expansion ratios, and the depth/width/resolution scaling were all co-optimized with SE attention present. Replacing SE with coordinate attention post-hoc, without re-running the architecture search, means the architecture is suboptimal for coordinate attention. The fact that coordinate attention still improves performance despite this mismatch is encouraging, but the true potential on NAS-derived architectures would require searching with coordinate attention from scratch. The paper treats the EfficientNet result as a strong demonstration, but it is more accurately a lower bound on coordinate attention's potential in searched architectures.

The Pascal VOC detection result where SE and CBAM provide zero gain is unexplained and warrants scrutiny. Table 7 shows SE and CBAM matching the vanilla MobileNetV2 at 71.7% mAP, while coordinate attention reaches 73.1% (+1.4). This is a striking result—SE attention provides gains on ImageNet, COCO, and both segmentation benchmarks, but fails completely on VOC detection. The paper does not discuss this anomaly. Possible explanations include: (1) the VOC detection fine-tuning protocol or hyperparameters were accidentally suboptimal for SE and CBAM; (2) the SSDLite320 detector on VOC has specific interactions with attention mechanisms that differ from COCO; (3) random variability (VOC 2007 test has 4,952 images, which should be large enough to suppress noise). Without multiple runs or investigation, this result should be treated cautiously—it may reflect a genuine limitation of channel-only attention for certain detection configurations, or it may be an experimental artifact.

Latency measurements are reported only for the ablation in Table 1. The main comparison tables (Tables 2–5) report parameters and M-Adds but not latency, and the latency numbers in Table 1 (17–19ms for coordinate attention vs. 16–18ms for SE attention on a Pixel 4) are not broken down by network width or architecture. For a paper targeting "efficient mobile network design," on-device latency is arguably more important than parameter count or theoretical M-Adds—mobile deployment is constrained by real-time inference speed and energy consumption, not just memory. A comprehensive latency comparison across all configurations, including different weight multipliers and both backbone families, would substantially strengthen the practical deployment case.

What experiment would most strengthen the paper? A direct measurement of the attention maps' spatial alignment with object positions. For example, on the COCO dataset, computing the correlation between the product attention map $g^h(i) \times g^w(j)$ (averaged across channels or for the most-attended channels) and the ground-truth bounding box masks would provide direct evidence that coordinate attention learns to spatially localize objects. If the attention maps systematically highlight object regions and suppress background, the "precise positional information" claim would be proven rather than inferred from task performance. If the attention maps look diffuse or uninterpretable despite improving accuracy, that would be equally informative—it would suggest the mechanism works through a different channel than explicit spatial localization.

6. Limitations and Trade-offs

6.1 The Difficulty Estimation Cost Is Unaccounted for in the Headline Efficiency Gains

The assumption or constraint. The entire compute-optimal framework depends on knowing the difficulty of each prompt before allocating the inference budget. The paper's method for estimating difficulty—generating 2048 samples per question and computing either ground-truth pass@1 (oracle) or the average PRM final-answer score (predicted)—is acknowledged in Section 3.2 as computationally expensive:

"estimating difficulty in this way still incurs additional computation cost during inference... our experiments do not account for this cost largely for simplicity"

The predicted difficulty variant still requires 2048 samples and PRM scoring per question, which for a 500-question test set means over 1 million extra generations and PRM forward passes before any test-time compute strategy is applied.

The consequence. The reported ~4× efficiency gains over best-of-N (Figures 4, 8) are computed after difficulty is known, with the estimation cost amortized across zero queries. In a realistic deployment, a system must pay the difficulty estimation cost for each prompt, and this cost can dominate the total compute—generating 2048 samples to estimate difficulty for a question that would otherwise receive 64 generations of test-time compute means the estimation phase consumes ~32× more compute than the actual problem-solving phase. The compute-optimal policy might still be valuable for batch processing of large question sets (where the 2048 samples can be generated once and reused), but for online or low-volume deployment, the overhead is prohibitive.

What evidence exists in the paper. The paper provides no experiment measuring total compute including difficulty estimation. The oracle and predicted difficulty curves in Figures 4 and 8 both assume difficulty is known without counting the cost of determining it. The paper does not report how much worse a strategy using zero difficulty samples (e.g., always applying the average-best strategy) would perform, nor does it measure how few samples are needed for a useful difficulty estimate—could 16 or 64 samples suffice? This is a gap in the experimental design.

Mitigation status. The authors explicitly flag this in Section 3.2 as a direction for future work: "pretraining or finetuning models to directly predict difficulty of a question" (Section 8). The current paper provides no amortization scheme, no lightweight difficulty estimator, and no analysis of how the $4 \times$ figure degrades when difficulty estimation cost is included. The $4 \times$ efficiency claim is therefore best understood as an upper bound contingent on solving the difficulty estimation problem separately.

6.2 Test-Time Compute Cannot Compensate for Fundamental Capability Gaps on Hard Problems

The assumption or constraint. The approach assumes the base model is capable of producing correct answers at some non-trivial rate—that correct solutions exist in the model's output distribution to be found through search or refinement. The paper states this boundary condition explicitly in the Section 7 takeaway:

"test-time compute amplifies existing capability but does not create it from nothing."

On the hardest difficulty quintile (bin 5), the base model's pass@1 is near zero, and no amount of search or revision produces meaningful improvement.

The consequence. For problems that genuinely exceed the base model's reasoning capabilities—novel mathematical derivations, out-of-distribution problem structures, complex multi-step reasoning requiring capabilities not acquired during pretraining—the compute-optimal framework offers essentially zero benefit. Figures 3 (right), 7 (right), and 9 all show bin 5 accuracy flatlining near 0–5% regardless of compute budget or strategy. An organization that deploys this system and encounters a stream of bin-5-difficulty prompts will see no improvement from test-time compute investment; pretraining a larger model or collecting additional training data remains the only viable path. This limits the approach's applicability to problems within the base model's "easy-hard" boundary—a boundary that is unknown a priori and varies across domains.

What evidence exists in the paper. The difficulty-bin breakdowns provide extensive evidence: Figure 3 (right) shows both best-of-N and beam search at ~1–3% accuracy on bin 5 across all budgets; Figure 7 (right) shows revision strategies at ~2–3% across all sequential-to-parallel ratios; Figure 9 (bottom lines) shows bin 5 compute-optimal scaling curves essentially flat and well below the larger model's performance. The FLOPs-matched comparison (Section 7) quantifies the worst-case disadvantage: at $R \gg 1$ with PRM search, hard problems show a −52.9% relative disadvantage from test-time compute compared to training a ~14× larger model. The evidence is consistent and unambiguous.

Mitigation status. The paper does not attempt to solve this limitation; it documents it with transparency. The authors position test-time compute as complementary to pretraining, not as a replacement for it, and the bin 5 results define the boundary between these complementary regimes. No mitigation is proposed beyond scaling pretraining for out-of-capability problems. This is a fundamental limitation rather than a fixable shortcoming of the method—it reflects the impossibility of sampling correct solutions that do not exist in the model's distribution.

6.3 The Revision Model Has a 38% Correct-to-Incorrect Reversion Rate

The assumption or constraint. The revision model is trained exclusively on trajectories where all in-context answers are incorrect followed by a correct target, using offline data construction with edit-distance-based pairing (Section 6.1). At inference time, the model produces sequential revisions—each conditioned on the previous answer—but may encounter correct answers in its history when an earlier revision happens to be correct. Since the model was never trained to handle correct preceding answers, it has no learned behavior for this situation.

The consequence. The paper reports in Section 6.1 that approximately 38% of correct answers in a revision chain get "revised" back to incorrect answers in the subsequent step. This means that even when the model produces a correct answer at step $k$ , there is a substantial probability that step $k+1$ will overwrite it with an incorrect one. Without mitigation, simply taking the final revision output would cap accuracy well below what the model can actually produce at some point in the chain—the revision trajectory is not monotonically improving.

What evidence exists in the paper. The 38% figure is cited in Section 6.1, though the paper does not provide a detailed breakdown (e.g., does the reversion rate vary across difficulty bins? Does it decrease as the chain progresses?). Figure 6 (left) shows that pass@1 at each step in the revision chain gradually improves but never exceeds ~25%, while the verifier-based selection across the chain achieves ~41.5%—confirming that within-chain selection is critical for recovering the correct answers that get "lost" to revisions. The paper does not ablate how the reversion rate changes with chain length.

Mitigation status. The paper partially mitigates this with within-chain selection (majority voting or verifier-based selection, described in Section 6.1): rather than taking only the final revision, the system evaluates all answers in the chain and selects the best one. This patches the symptom but does not address the root cause—the model's training distribution mismatch. The paper does not explore training the revision model on trajectories that include correct answers (learning to recognize when no revision is needed—a "stop revising" signal), which would be the principled solution. The ReST $^{EM}$ experiment (Appendix K, Figure 16) further highlights the fragility: optimizing the revision model with on-policy data caused performance to degrade, suggesting revision training is highly sensitive to methodology.

6.4 Single Benchmark, Single Model Family Evaluated

The assumption or constraint. All experiments use the MATH benchmark (500 test questions) with PaLM 2-S* as the base model. The paper provides no results on other benchmarks, other model families, or other reasoning domains. The authors state their belief that PaLM 2-S* is "representative of the capabilities of many contemporary LLMs" (Section 4), but this is an assertion without empirical backing.

The consequence. Several findings could be model- or domain-specific. The PRM's quality and over-optimization behavior (the threshold where beam search degrades performance on easy problems, shown in Figure 3 right) depend on PaLM 2-S*'s output distribution—a different model with better calibration might exhibit different over-optimization patterns, changing which strategies are optimal at which difficulty levels. The revision model's ability to learn from edit-distance-paired trajectories depends on the base model's in-context learning and instruction-following capabilities, which vary substantially across model families. The MATH benchmark tests competition-level symbolic math reasoning—the difficulty-dependent patterns (beam search hurting easy problems, revisions helping easy problems, the sharp distinction between bins 3 and 4) might not generalize to code generation, logical reasoning, scientific QA, or tasks requiring factual knowledge rather than inference. The test set size of 500 questions, split into five quintiles of ~100 each, means strategy selection within each difficulty bin is based on only ~50 questions per cross-validation fold—a sample size that introduces non-trivial variance in the computed-optimal policy.

What evidence exists in the paper. None—this is a scope limitation. All claims are specific to MATH with PaLM 2-S*. The paper does not include results on GSM8K (another math benchmark that would test generalization to different difficulty distributions), HumanEval (code generation), or any non-math reasoning task. The compute-optimal strategies selected for MATH might not transfer to other benchmarks, and the paper provides no evidence to the contrary.

Mitigation status. Not addressed. The authors acknowledge this implicitly by not making claims beyond their experimental scope, but they do not discuss it as an explicit limitation. Replication on at minimum one additional benchmark and one additional model family would substantially strengthen confidence in the generalizability of the difficulty-conditioned scaling framework.

6.5 Revisions and Search Are Studied Independently; the Combined System Is Not Evaluated

The assumption or constraint. The paper studies two complementary test-time compute mechanisms—PRM-guided search (Section 5) and iterative revisions (Section 6)—but evaluates them as independent pipelines. The compute-optimal strategy selects between search algorithms and revision ratios separately; there is no joint strategy that combines PRM tree-search with the revision model.

The consequence. The current results likely represent a lower bound on what a fully integrated system could achieve. The two mechanisms have complementary strengths: revisions improve the proposal distribution (generating better candidate solutions by conditioning on past mistakes), while PRM search improves selection (identifying the best among generated candidates). A combined system could, for example, use the revision model as the proposal distribution within beam search—at each expansion step, the model conditions on previous rejected branches to produce higher-quality candidates—or use the PRM to guide which revisions to pursue, pruning unpromising revision chains early. The paper's difficulty-dependent analysis suggests these mechanisms work best on different problem types: revisions dominate on easy problems, beam search dominates on medium problems, and a combined approach might outperform either alone on medium-hard problems where multiple high-level approaches (parallel search) combined with local refinement (revisions) could be optimal.

What evidence exists in the paper. Section 8 explicitly acknowledges this gap: "we did not experiment with PRM tree-search techniques in combination with revisions." No ablation studies the interaction, no results are provided, and the paper's main compute-optimal curves (Figures 4, 8) treat search and revisions as alternatives rather than components of a unified strategy.

Mitigation status. Acknowledged as future work (Section 8). This is the most obvious next step from the paper's framework, and the absence of combined results means the paper's headline efficiency gains may understate what the full framework can achieve. The paper's contributions—the difficulty-conditioned allocation principle and the characterization of when search vs. revisions work best—provide clear guidance for designing such a combined system, but do not themselves demonstrate it.

6.6 No Accounting for Latency or Wall-Clock Time in the Compute-Optimal Strategy

The assumption or constraint. The paper measures compute exclusively in "generations" (number of complete solutions sampled), which is a reasonable proxy for total FLOPs but ignores wall-clock time and the serial vs. parallel execution characteristics of different strategies. The compute-optimal policy freely selects strategies with different latency profiles—for example, a 128-generation budget allocated as 64 sequential revisions × 2 parallel chains requires roughly 64× the wall-clock time of 128 fully parallel samples, assuming sufficient hardware parallelism for the latter.

The consequence. For latency-sensitive applications—interactive assistants, real-time tutoring, on-device deployment where users expect sub-second responses—the sequential-heavy strategies favored by the compute-optimal policy on easy problems may be impractical regardless of their accuracy advantages. A 64-step revision chain on a modest GPU might take tens of seconds, while 128 parallel samples could complete in the time of a single generation. The paper's claim that compute-optimal scaling provides "more than 4× better efficiency" (Abstract) would not hold under a latency budget—what matters is not total FLOPs but time-to-answer. An organization deploying this in production would need to constrain the sequential-to-parallel ratio by a maximum acceptable latency, which could significantly reduce the achievable efficiency gains on easy problems (where the optimal strategy is often purely sequential).

What evidence exists in the paper. None. The paper does not report wall-clock time for any configuration, does not discuss latency constraints, and does not evaluate how the compute-optimal policy would change under a latency budget. The parallel baseline (best-of-N) is naturally suited to low-latency deployment since all samples can be generated simultaneously; the sequential revision strategies are not. This is a practical tradeoff the paper does not analyze, and it is particularly relevant given that mobile deployment (low-latency on-device inference) is one of the motivating applications (Section 1).

Mitigation status. Not addressed. The paper's generation-count metric is standard in the LLM scaling literature and is appropriate for the paper's primary contribution (a scaling analysis analogous to Chinchilla), but the transition from research analysis to production deployment requires latency considerations that the paper does not provide. Future work could optimize for accuracy under a joint compute-latency budget, which would likely shift the optimal sequential-to-parallel ratios toward more parallel configurations, especially at low latency targets.

7. Implications and Future Directions

How This Work Changes the Landscape

Coordinate attention introduces a third design paradigm for efficient spatial attention in mobile networks, positioned between the two extremes that previously defined the field. Before this work, the implicit consensus was that spatial attention for mobile networks either had to be local (captured by convolutions with fixed kernel sizes, as in CBAM's 7×7 convolution) or global-and-expensive (captured by non-local self-attention with quadratic cost). The first option is computationally cheap but spatially myopic—a single convolution layer cannot relate distant pixels. The second option captures true long-range dependencies but is, as the paper states explicitly, "not suitable for mobile networks" due to the cost of computing pairwise relationships between all spatial positions. The paper's core conceptual move is demonstrating that global context along one spatial direction combined with precise positional preservation along the orthogonal direction—achieved by factorizing 2D pooling into two 1D encoding processes—captures enough spatial structure to substantially improve performance on spatially-demanding tasks while remaining computationally viable for mobile deployment.

This is a reframing rather than a paradigm shift. The underlying operations (1D average pooling, 1×1 convolutions, sigmoid gating) are all standard. What changes is the realization that spatial attention need not be isotropic—it need not treat all spatial directions equally or relate all pairs of positions. By decomposing spatial context into separable horizontal and vertical components and recombining them multiplicatively, the network can attend to specific coordinate regions (rows × columns) without ever computing or storing a full H×W attention matrix. The computational cost scales as O(C×(H+W))—linear in the spatial dimensions—rather than the O(H²W²) of pairwise self-attention or even the O(k²×H×W) of large-kernel convolutions. For mobile networks where every FLOP matters, opening this third path between "too local" and "too expensive" is a genuine conceptual contribution, even though the architectural components are individually simple.

The paper also introduces a diagnostic framework that will likely influence how future attention mechanisms are evaluated. By testing across a spectrum of tasks with increasing spatial precision demands—ImageNet classification (scalar output per image), COCO object detection (bounding boxes), Pascal VOC and Cityscapes semantic segmentation (per-pixel labels)—the paper demonstrates that the performance gap between different attention mechanisms is not uniform across tasks but reflects what kind of information each mechanism encodes. SE attention, which mathematically discards all positional information through 2D global pooling (Eqn. 1), shows its strongest relative performance on ImageNet classification (+1.2 over baseline) but provides diminishing returns on dense prediction tasks (only +0.85 mIoU on Pascal VOC segmentation at stride 16). CBAM, which attempts spatial attention through a local 7×7 convolution on aggressively channel-squeezed features, actually degrades to baseline-level performance on Pascal VOC detection (71.7% mAP, identical to no-attention) and Cityscapes segmentation (71.4% mIoU, identical to no-attention). Coordinate attention, which explicitly encodes positional information, shows its largest gains on segmentation (+2.48 mIoU over baseline on Pascal VOC, +2.6 on Cityscapes)—precisely where spatial precision matters most. This pattern is not merely an empirical curiosity; it validates the paper's mechanistic analysis and provides a transferable evaluation methodology. The field now has a principled reason to evaluate attention mechanisms on dense prediction tasks, not just ImageNet top-1, because the shape of the performance curve across tasks reveals what the attention mechanism actually learns to encode.

The work also reconciles a latent tension in the mobile network literature. SE attention had been the dominant attention mechanism for mobile networks since its introduction, achieving consistent gains on ImageNet classification that made it a standard component in architectures ranging from MobileNetV3 to EfficientNet. CBAM was proposed as an improvement by adding spatial attention, but its benefits on mobile-scale networks were unclear—the paper's experiments across three tasks and multiple architectures demonstrate that CBAM's spatial attention module, far from helping, often hurts performance on dense prediction tasks (Pascal VOC detection: 71.7% for CBAM vs. 71.7% for no-attention; Cityscapes segmentation: 71.4% for CBAM vs. 71.4% for no-attention). This resolves the question of whether CBAM's approach to spatial attention (channel squeezing to 2 channels + 7×7 convolution) is appropriate for mobile networks—it is not, and the paper provides mechanistic explanations (information loss from channel squeezing, local-only receptive field) that explain the failure. This makes the design space clearer: SE attention defines the floor for what channel-only attention can achieve; coordinate attention defines a new, higher ceiling for attention that also encodes spatial structure; and the specific design choices in CBAM (aggressive channel reduction, local convolution) represent a path that should be avoided in future mobile attention design.

Finally, the paper shifts attention research away from exclusively pursuing more complex attention operations and toward rethinking how spatial context is aggregated. The past several years saw a proliferation of increasingly sophisticated attention mechanisms—non-local blocks, criss-cross attention, dual attention, self-calibrated convolutions, triplet attention—each adding more operations and parameters to capture spatial relationships. Coordinate attention achieves its gains with operations no more complex than 1D average pooling and 1×1 convolutions, suggesting that the key insight is not engineering a more powerful attention operator but rather choosing the right spatial aggregation pattern for the computational budget. This is a different kind of research contribution than a new attention formula—it is a design philosophy that may influence how future work approaches the efficiency-expressiveness tradeoff.

However, the paper's impact has clear boundaries. It does not replace self-attention for tasks where per-pair spatial relationships are essential (fine-grained correspondence, complex non-axis-aligned spatial reasoning). It does not address temporal attention or cross-modal attention. And its central mechanism—factorization into orthogonal 1D encodings with multiplicative recombination—assumes that spatial attention patterns are approximately separable into row-wise and column-wise components. For visual tasks where this separability assumption breaks down (diagonal structures, objects with complex non-rectangular spatial extent, scenes where "important rows" and "important columns" do not cleanly intersect at object locations), the approach may underperform relative to mechanisms that can model non-separable spatial attention. The paper provides no negative results on such cases, which limits our understanding of the boundary conditions.

Follow-Up Research This Work Enables

Direct comparison with axial attention under identical mobile budgets. The paper motivates coordinate attention partly in contrast to non-local self-attention, arguing that full pairwise attention is too expensive for mobile networks. However, axial attention—which factorizes self-attention along height and width dimensions, computing attention separately along each axis and combining the results—operates on the same dimensional factorization principle as coordinate attention but with a learned attention mechanism rather than global average pooling. The paper does not compare against axial attention, which is a significant omission given the conceptual similarity. A strong follow-up would implement axial attention blocks with the same parameter and FLOP budgets as the coordinate attention blocks in Tables 2–5, evaluating on the identical suite of tasks (ImageNet classification, COCO detection, Pascal VOC and Cityscapes segmentation). The specific question is: does learned axial self-attention (where the attention weights are dynamically computed from feature similarities) outperform static 1D global pooling (where the "attention" is just average pooling followed by learned channel gating) under a tight mobile compute budget? If axial attention wins, it suggests that learned spatial attention weights capture information that global average pooling misses, even along a single dimension. If coordinate attention wins, it suggests that the inductive bias of global average pooling (equal weighting of all positions along the pooled dimension) is a beneficial regularizer that learned attention, with its additional parameters, cannot overcome given limited mobile-scale training data. The Cityscapes segmentation benchmark (Table 9) would be the most diagnostic test case, since segmentation requires the finest spatial reasoning and would most clearly reveal whether learned axial attention provides localization benefits over global pooling.

Measuring the spatial localization quality of the learned attention maps directly. The paper claims that coordinate attention captures "precise positional information" and helps the network "more accurately locate the objects of interest," but this claim is supported only indirectly through improved task performance and Grad-CAM visualizations (Figure 4). A rigorous follow-up would directly measure the spatial alignment between the coordinate attention maps and object positions. Concretely: on the COCO dataset, for each image, extract the per-channel attention weights g^h(i) and g^w(j) from the coordinate attention block at a specific layer. Compute the outer product g^h(i) × g^w(j) to get a 2D attention map per channel. Aggregate these across channels (e.g., by averaging all channels or selecting the top-k channels by mean activation). Measure the Intersection-over-Union (IoU) between the binarized attention map (thresholded at, say, 0.5) and the ground-truth object bounding boxes or segmentation masks. Compare this IoU against the same metric computed from SE attention (which only has channel weights and therefore a spatially uniform attention map) and CBAM (which has a single spatial attention map shared across channels). If coordinate attention achieves substantially higher IoU, the "precise positional information" claim would be directly validated. If the IoU is low (the attention maps are diffuse or uninterpretable despite improving task performance), that would be equally informative—it would indicate that coordinate attention improves performance through a mechanism other than explicit spatial localization (perhaps through improved channel-wise feature recalibration that implicitly benefits spatial tasks), which would reframe our understanding of what the mechanism actually does. This experiment would also reveal whether the multiplicative gating g^h(i) × g^w(j) produces attention maps that align with entire objects, object parts, or some other spatial pattern that does not correspond to human-interpretable object locations.

Stress-testing the axis-aligned separability assumption on rotated or diagonally-structured visual data. The coordinate attention mechanism assumes that spatial attention patterns are approximately separable into row-wise and column-wise components—the attention weight at position (i,j) is the product of a row-dependent term and a column-dependent term. This assumption is reasonable for many natural images where objects have approximately axis-aligned spatial extent (faces are roughly centered and upright, buildings have horizontal and vertical edges). However, it will fail on visual data where the relevant spatial patterns are fundamentally non-separable: diagonal structures, rotated objects, or scenes where the "important rows" and "important columns" do not form a rectangular region around the object of interest. A stress-test follow-up would construct or identify datasets where this failure mode is likely, and compare coordinate attention against SE attention and a small non-local attention baseline on those datasets. Candidates include: (1) a rotated version of ImageNet or COCO where all images are rotated by 45 degrees (diagonal edges and object boundaries that are not axis-aligned); (2) the DOTA dataset for aerial object detection, where objects appear at arbitrary orientations and are often densely packed; (3) a synthetic dataset of diagonal line segments or curves where the ground-truth "attention" pattern is known to be non-separable. If coordinate attention maintains its advantage over SE attention on rotated data, it suggests the mechanism is more robust than the separability assumption would predict (perhaps the network learns rotation-invariant features in earlier layers before the attention block, or the multiplicative gating captures non-axis-aligned patterns indirectly). If coordinate attention degrades to SE-level performance on rotated data while a non-local attention baseline maintains its advantage, it would establish a clear boundary condition: coordinate attention is appropriate for tasks where objects are predominantly axis-aligned, and should be replaced with something more flexible when arbitrary orientations are expected. This boundary condition would be practically important for applications like aerial imagery, medical imaging (where anatomy can appear at any orientation), or robotics (where the camera may not be upright).

Architecture search with coordinate attention as a first-class component. The EfficientNet-b0 result in Table 5 is promising but limited: the paper simply replaces SE attention with coordinate attention in an architecture that was searched with SE attention in place, producing a +0.6 top-1 improvement. The true potential of coordinate attention in NAS-derived architectures would be revealed by running a full architecture search with coordinate attention blocks as an available operation from the start, allowing the search algorithm to learn optimal insertion positions, reduction ratios, and interactions with other architectural choices (kernel sizes, expansion ratios, depth/width scaling). A follow-up could integrate coordinate attention into a mobile-scale NAS framework (e.g., ProxylessNAS, MnasNet, or TuNAS) as a drop-in replacement for SE attention, then re-run the search with the same search space and constraints. The primary metric would be the ImageNet top-1 accuracy of the searched coordinate-attention architecture compared to the original SE-attention architecture at the same parameter/FLOP count, plus transfer performance on COCO detection and Cityscapes segmentation. If the searched coordinate-attention architecture outperforms the best SE-attention architecture by a margin larger than the +0.6 seen from simple replacement in EfficientNet-b0, it would demonstrate that the NAS process can exploit coordinate attention's spatial encoding in ways that manual insertion cannot—for example, by learning to place coordinate attention blocks selectively at layers where spatial information is most valuable, and using SE attention (or no attention) elsewhere. This would also test whether coordinate attention's larger bottleneck parameter count (the two independent 1×1 convolutions F_h and F_w) trades off against other parameter uses in the architecture search, or whether the spatial information benefit outweighs the parameter cost regardless.

Training a lightweight difficulty predictor from the PRM score distribution to close the deployment overhead gap. While not part of this paper (which focuses exclusively on vision), the broader framework this paper exemplifies—replacing an expensive 2D operation (2D global pooling) with two cheaper 1D operations to preserve structure—suggests an analogous approach for the difficulty estimation problem in test-time compute scaling. A follow-up could explore whether the PRM's per-step score distribution can be used to predict question difficulty from a small number of initial samples (e.g., 8–16 rather than 2048), analogous to how coordinate attention preserves positional information by not collapsing both spatial dimensions simultaneously. Concretely: generate K samples (where K is small, e.g., 8) from the base model, compute per-step PRM scores for each sample, and extract features from the score distribution (mean, variance, minimum score across steps, trajectory of scores within each solution). Train a lightweight classifier (a small MLP or even logistic regression) to predict the oracle difficulty bin from these distributional features. Evaluate whether 8-sample predicted difficulty achieves comparable compute-optimal scaling curves to the full 2048-sample predicted difficulty (Figure 4). If successful, this would make the compute-optimal framework deployable without the prohibitive difficulty estimation overhead, closing the most significant practical gap the paper acknowledges.

Ablating the importance of the shared 1×1 convolution and the split-then-independent-transform design. The coordinate attention generation stage (Section 3.2.2) concatenates the horizontal and vertical feature maps, passes them through a shared 1×1 convolution F_1, splits the result, and then applies two independent 1×1 convolutions F_h and F_w. This design has a specific rationale: the shared transformation forces a common encoding of spatial-direction information, while the independent transformations allow direction-specific attention weight prediction. However, the paper never ablates whether this specific architecture matters. A follow-up could test three variants: (1) fully shared: the same 1×1 convolution is used for both horizontal and vertical branches at every stage (no split, no independent F_h and F_w)—this would test whether direction-specificity is necessary or whether a common transformation suffices; (2) fully independent: no shared F_1 at all—the horizontal and vertical features are processed by entirely separate 1×1 convolutions from the start; (3) no bottleneck: the concatenated features are projected directly to C channels with a single 1×1 convolution (no reduction ratio r), then split and sigmoid-activated. These ablations would reveal which aspect of the design contributes to performance: the parameter efficiency of sharing, the expressiveness of direction-specificity, or the regularizing effect of the bottleneck reduction. A negative result—e.g., fully independent transformations matching or exceeding the shared-then-split design—would simplify the architecture and reduce parameters without losing accuracy. A positive result—the shared-then-split design outperforming both alternatives—would validate the paper's specific architectural choice and provide a design principle (shared encoding of spatial directions followed by direction-specific decoding) that could guide future attention mechanisms.

Evaluating coordinate attention on video understanding tasks where spatiotemporal factorization is natural. The paper evaluates on static image tasks, but the 1D factorization idea extends naturally to video, where an analogous design could factor 3D spatiotemporal pooling (which would collapse H, W, and T simultaneously) into three separate 1D encodings: height, width, and time. A follow-up could implement a "spatiotemporal coordinate attention" block for mobile video networks (e.g., MobileNetV2-3D or a lightweight (2+1)D convolution architecture), with three parallel 1D pooling operations: horizontal spatial pooling (producing C×T×H×1 features), vertical spatial pooling (C×T×1×W), and temporal pooling (C×1×H×W). The concatenation, shared 1×1 convolution, split, and independent encoding would extend naturally to three branches. This could be evaluated on action recognition (Something-Something V2, which requires fine-grained temporal reasoning, or Kinetics-400 for more static appearance-based recognition) and compared against SE attention (3D global pooling → channel attention) and CBAM-style spatial attention with 3D convolutions. The hypothesis: the spatiotemporal coordinate attention would show larger gains on temporally-demanding tasks (Something-Something, where knowing when motion occurs matters) than on appearance-dominated tasks (Kinetics, where static frame-level features may suffice), analogous to how the original coordinate attention shows larger gains on spatially-demanding segmentation tasks than on classification.

Practical Applications and Downstream Use Cases

On-device semantic segmentation for mobile photography and augmented reality. The Cityscapes result in Table 9—74.0% mIoU with coordinate attention vs. 71.4% for the vanilla MobileNetV2 backbone, at output stride 8 with full 1024×2048 resolution testing—directly translates to better segmentation quality on mobile devices. In mobile photography, semantic segmentation enables portrait mode (separating foreground subjects from backgrounds for synthetic bokeh), sky replacement, and selective image enhancement (brightening faces without affecting backgrounds). A +2.6 mIoU improvement on a Cityscapes-scale task, achieved with only 0.5M additional parameters (5.0M vs. 4.5M) and running within the latency constraints implied by the Pixel 4 measurements (17–19ms vs. 14–16ms baseline), means that phone manufacturers could deploy better segmentation models without increasing the model download size or inference time beyond what current hardware supports. The Pascal VOC segmentation results (+2.5 mIoU at stride 16, +2.1 mIoU at stride 8) suggest the benefit generalizes across segmentation domains, not just street scenes. For augmented reality applications—where a mobile device must segment surfaces, objects, and people in real-time to place virtual content correctly—the improved spatial localization that coordinate attention provides (demonstrated qualitatively in the Grad-CAM visualizations of Figure 4) could reduce boundary artifacts and mislocalization of virtual objects.

Mobile object detection for retail, inventory, and assistive technology. The COCO detection results in Table 6 (24.5 AP for coordinate attention vs. 23.7 AP for SE attention with SSDLite320) and especially the Pascal VOC result in Table 7 (73.1% mAP for coordinate attention vs. 71.7% for both SE and CBAM) demonstrate that coordinate attention improves mobile object detection, sometimes in regimes where other attention mechanisms provide zero benefit. For retail inventory systems running on handheld devices (barcode scanning, shelf auditing, product recognition), the improved detection AP across all object scales (Table 6: APS 2.3 vs. 2.2, APM 26.2 vs. 25.4, APL 45.9 vs. 44.7 for coordinate attention vs. SE) means more reliable detection of both small products (individual items on a shelf) and large ones (display cases). For assistive technology applications—phone apps that help visually impaired users identify objects in their environment—the improved spatial localization (the core claimed benefit of coordinate attention) could help the system describe not just what objects are present but where they are relative to the user, which requires the precise positional information that coordinate attention is designed to preserve. The fact that coordinate attention's detection gains are consistent across both COCO (80 object categories, complex scenes) and Pascal VOC (20 categories, cleaner compositions) suggests the benefit is not domain-specific.

Drop-in replacement for SE attention in existing mobile architectures with minimal engineering effort. The paper's integration pattern (Figure 3) demonstrates that coordinate attention can replace SE attention at identical insertion points in MobileNetV2, MobileNeXt, and EfficientNet with no changes to the backbone architecture. This is a practical advantage that accelerates adoption: teams using these mobile architectures can replace SE attention blocks with coordinate attention blocks by swapping the attention module implementation, without redesigning the network, changing training hyperparameters, or modifying the deployment pipeline. The consistent gains across weight multipliers (Tables 2 and 3: +0.5 to +0.8 top-1 across 0.5×, 0.75×, and 1.0× widths) mean that the improvement holds regardless of which model size variant a team is using, from the smallest (2.0M parameters) to the largest (3.5M+). The latency increase of 1–3ms on a Pixel 4 (Table 1) sets a concrete expectation for the engineering cost: applications with latency budgets above ~20ms per inference can adopt coordinate attention with negligible user-facing impact, while those with tighter budgets (e.g., real-time video processing at 30+ FPS requiring <30ms per frame) may need to evaluate whether the accuracy gain justifies the latency cost.

When to Prefer This Method

The paper does not explicitly articulate a decision framework for choosing between coordinate attention and the alternatives it evaluates (SE attention, CBAM). However, its experimental results across tasks, architectures, and model scales implicitly define the conditions under which coordinate attention should be preferred. These conditions are derived from the paper's evidence rather than stated by the authors as explicit guidance:

Prefer coordinate attention over SE attention when dense spatial prediction tasks (detection, segmentation) are in the deployment pipeline, not just classification. The gap between coordinate attention and SE attention grows substantially on segmentation (Pascal VOC: +1.6 mIoU at stride 16; Cityscapes: +1.8 mIoU) and detection (COCO: +0.8 AP; Pascal VOC: +1.4 mAP, where SE provides zero gain) compared to classification (+0.5 to +0.8 top-1). If the backbone will be fine-tuned for detection or segmentation—the most common production pipeline for mobile vision—coordinate attention's advantages are largest precisely where they matter most.
Prefer SE attention when implementation simplicity is paramount and classification is the primary or only task. SE attention is simpler (one global pooling, two fully-connected layers), has a longer track record of deployment, and achieves close to coordinate attention's performance on ImageNet classification (within 0.5–0.8 top-1). If the use case is pure image classification (e.g., photo organization, content moderation where only image-level labels are needed), the additional complexity of coordinate attention's two-branch pooling, concatenation, split, and independent transformations may not justify a fraction of a percentage point on top-1 accuracy.
Prefer coordinate attention over CBAM in all mobile settings. The paper shows CBAM matching or underperforming SE attention on every benchmark, and completely failing (zero improvement over no-attention) on Pascal VOC detection and Cityscapes segmentation. There is no experimental regime in this paper where CBAM outperforms coordinate attention or even convincingly outperforms SE attention. For mobile networks, CBAM's spatial attention design (channel squeezing to 2, 7×7 convolution) appears to be counterproductive, and coordinate attention provides a strictly superior alternative for encoding spatial information.
Prefer coordinate attention when deploying on hardware that can tolerate a 6–19% latency increase over the SE-attention baseline. The Pixel 4 latency measurements in Table 1 (17–19ms for coordinate attention vs. 16–18ms for SE attention vs. 14–16ms for no attention) provide concrete guidance: the additional cost of coordinate attention over SE attention is 1–3ms, representing a 6–19% relative increase. Applications with headroom in their latency budget (e.g., single-image processing where 20ms vs. 17ms is imperceptible) should adopt coordinate attention for the dense-prediction gains. Applications operating at the edge of real-time constraints (e.g., 60 FPS video processing where each frame has a 16.7ms budget) may need to benchmark on their specific hardware, since the absolute 17–19ms range already exceeds a 60 FPS budget regardless of attention choice.
Prefer coordinate attention at all model scales within the tested range (2.0M–6.1M parameters). The consistent gains across MobileNetV2 weight multipliers (0.5×, 0.75×, 1.0× in Table 2), MobileNeXt weight multipliers (Table 3), and EfficientNet-b0 (Table 5) demonstrate that the relative benefit does not diminish at smaller model sizes. The 0.5× MobileNetV2 with coordinate attention (67.0% top-1) outperforms the 0.75× version with SE attention (71.5%—though this crosses weight multipliers rather than being a like-for-like comparison). This suggests coordinate attention is not only for the largest mobile models; it improves even the smallest variants proportionally.