Coordinate Attention for Efficient Mobile Network Design

The paper addresses a limitation of channel attention mechanisms like Squeeze-and-Excitation in mobile network design: transforming a feature tensor to a single vector via 2D global pooling loses positional cues. The authors propose coordinate attention, which factorizes channel attention into two 1D feature encoding processes that aggregate features separately along the two spatial directions. This produces a pair of direction-aware, position-sensitive attention maps that capture long-range dependencies along one axis while retaining precise positional information along the other, applied complementarily to the input features.

Coordinate attention is simple and can be flexibly inserted into classic mobile architectures such as MobileNetV2, MobileNeXt and EfficientNet with nearly no computational overhead. Extensive experiments show it not only benefits ImageNet classification but behaves even better on downstream tasks like object detection and semantic segmentation, making it a practical drop-in module for efficient vision models.