DecTrans: Person Re-identification with Multifaceted Part Features via Decomposed Transformer

Utilizing part-level features provides a more detailed representation, leading to improved results in person re-identification (ReID). Yet existing works either use external tasks like pose estimation or struggle to define part features, which limit the model’s learning capability. In this work, we propose the Decomposed Transformer (DecTrans), a transformer-based person ReID framework which exploits multifaceted part features. In particular, DecTrans extracts local features using the Vision Transformer (ViT) and then maps them into latent parts through a novel Token Decomposition (TD) layer. In the TD layer, soft clustering of ViT tokens forms clusters, and each token is decomposed into components based on its similarity to all cluster centroids. Token components referencing the same cluster are then regrouped to produce part features, thereby retaining more feature details. To ensure tokens from different pedestrians but referring to the same part are sufficiently clustered together, we propose to remove id information from tokens before clustering. Besides, we also propose a simple yet efficient data augmentation named Image Graying, which has been experimentally validated when used in conjunction with the TD layer. The DecTrans achieves remarkable performance, e.g., mAP and Rank1 of 70.8 % & 87.1 %, and 61.6 % & 67.7 % on MSMT17 and Occluded-Duke, significantly outperforming state-of-the-arts.