- [pdf] [supp] [code]
Co-Attention Aligned Mutual Cross-Attention for Cloth-Changing Person Re-Identification
Person re-identification (Re-ID) has been widely studied and achieved significant progress. However, traditional person Re-ID methods primarily rely on cloth-related color appearance, which is unreliable under real-world scenarios when people change their clothes. Cloth-changing person Re-ID that takes this problem into account has received increasing attention recently, but it is more challenging to learn discriminative person identity features, since larger intra-class variation and smaller inter-class easily occur in the image feature space with clothing changes. Beyond appearance features, some known identity-related features can be implicitly encoded in images (e.g., body shapes). In this paper, we first design a novel Shape Semantics Embedding (SSE) module to encode body shape semantic information, which is one of the essential clues to distinguish pedestrians when their clothes change. To better complement image features, we further propose a Co-attention Aligned Mutual Cross-attention (CAMC) framework. Different from previous attention-based fusion strategies, it first aligns features from multiple modalities, then effectively interacts and transfers identity-aware but cloth-irrelevant knowledge between the image space and the body shape space, resulting in a more robust feature representation. To the best of our knowledge, this is the first work to adopt Transformer to handle the multi-modal interaction for cloth-changing person Re-ID. Extensive experiments demonstrate the effectiveness of our proposed method and show the superior performance achieved on several cloth-changing person Re-ID benchmarks. Codes will be available at https://github.com/QizaoWang/CAMC-CCReID.