Dual Attention Network for Scene Segmentation

This paper tackles scene segmentation by capturing rich contextual dependencies through a self-attention mechanism. Unlike prior work that captures context by fusing multi-scale features, the authors propose a Dual Attention Network (DANet) that adaptively integrates local features with their global dependencies. Two attention modules are appended on top of a traditional dilated fully convolutional network: a position attention module that aggregates the feature at each position as a weighted sum of features at all positions, and a channel attention module that emphasizes interdependent channel maps by integrating associated features across channels.

Summing the outputs of the two attention modules further improves feature representation and leads to more precise segmentation. DANet achieves new state-of-the-art performance on three challenging scene segmentation datasets, Cityscapes, PASCAL Context and COCO Stuff. In particular, it reaches a mean IoU of 81.5% on the Cityscapes test set without using coarse training data, demonstrating the effectiveness of jointly modeling spatial and channel-wise dependencies.