Enhance Heads in Vision Transformer for Occluded Person Re-Identification

Abstract

Occlusion scenarios pose a great challenge to person re-identification (ReID) task because various occlusions may weaken the discriminative features and introduce interference. Recently, transformer-based networks, which can aggregate features of all the image patches to construct global features adaptively, have shown advantages in occluded person ReID. Existing methods mainly adopted transformer as a feature extractor and enhanced local features from the output of the transformer encoder. However, during the processing of self-attention blocks, disturbing features from occlusions may be diffused into all the tokens, making it difficult to enhance local features effectively. On the other hand, the different heads in self-attention remain isolated during image encoding. Therefore, we consider applying feature enhancement strategies in the channel dimensions instead of the spatial dimensions. First, we divide the heads into groups to enhance diversity and strengthen the robustness of some patterns in occlusion scenarios. Then during training we iteratively suppress the most salient patterns, forcing the model to mine more salient patterns. Finally, we assign adaptive weights for different head groups to compute a robust distance matrix. Our method enhances the model’s ability to extract discriminative and diverse head features and achieves the state-of-the-art performance on occluded person ReID benchmarks, e.g., Rank-1 of 73.2% on Occluded-DukeMTMC.