Learning Disentangled Attribute Representations for Robust Pedestrian Attribute Recognition

Although various methods have been proposed for pedestrian attribute recognition, most studies follow the same feature learning mechanism, \ie, a shared image to classify multiple attributes. However, this mechanism leads low-confidence predictions and non-robustness of model in inference stage. In paper, we investigate why is case. We mathematically discover that central cause optimal cannot maintain high similarities with classifiers simultaneously context minimizing classification loss. addition, ignores spatial semantic distinctions between different To address these limitations, propose novel disentangled (DAFL) framework learn each attribute, which exploits characteristics The mainly consists learnable queries, cascaded semantic-spatial cross-attention (SSCA) module, group attention merging (GAM) module. Specifically, based on SSCA module iteratively enhances localization attribute-related regions aggregates region features into features, used updating queries. GAM splits attributes groups distribution utilizes reliable supervise query maps. Experiments PETA, RAPv1, PA100k, RAPv2 show method performs favorably against state-of-the-art methods.