In person re-identification, extracting part-level features from person images has been verified to be crucial. Most of existing CNN-based methods only locate the human parts coarsely, or rely on pre-trained human parsing models and fail in locating the identifiable non-human parts (e.g., knapsack). In this paper, we introduce an alignment scheme in Transformer architecture for the first time and propose the Auto-Aligned Transformer (AAformer) to automatically locate both the human parts and non-human ones at patch-level. We introduce the part tokens, which are learnable vectors, to extract part features in Transformer. A part token only interacts with a local subset of patches in self-attention and learns to be the part representation. To adaptively group the image patches into different subsets, we design the Auto-Alignment. Auto-Alignment employs a fast variant of Optimal Transport algorithm to online cluster the patch embeddings into several groups with the part tokens as their prototypes. We harmoniously integrate the part alignment into the self-attention and the output part tokens can be directly used for retrieval. Extensive experiments validate the effectiveness of part tokens and the superiority of AAformer over various state-of-the-art methods.
Download