Image-Attribute and Frequency-Spatial Dual Collaborative Learning for Pedestrian Attribute Recognition

The key of pedestrian attribute recognition (PAR) task lies in accurately extracting various attributes from pedestrian images, such as gender, age, clothing, and accessories. Prior CLIP-based PAR methods primarily rely on static textual prompts or template-based sentences and perform direct classification or one-to-one image-text matching. Although these methods have achieved promising results, their reliance on such static prompts may lack flexibility in capturing the dynamic interactions between multiple attribute labels and their fine-grained visual semantics. To address this, we propose a novel Frequency-Spatial and Image-Attribute (FSIA) method that adopts a dual collaborative learning strategy to effectively model image-attribute associations and enhance feature representation. FSIA consists of two key components: 1) An Image-Attribute Collaborative Learning (IACL) framework that integrates visual data with attribute labels to enable a nuanced semantic understanding. The proposed IACL utilizes learnable attribute prompts, each specifically optimized for an individual attribute category, facilitating expressive and discriminative visual-language alignment; 2) A Frequency-Spatial Collaborative Learning (FSCL) module that leverages frequency-domain information (often overlooked in prior PAR) to exploit dual-domain frequency-spatial information in pedestrian images, thereby enhancing feature robustness. Extensive experiments show that FSIA has significant advantages in improving the performance of PAR and the challenging zero-shot PAR task. Specifically, the proposed FSIA achieves 90.08% in mA, 82.09% in Accu, 87.92% in Prec, 89.73% in Recall and 88.45% in F1 on the PETA dataset.