ASRL: Correlation-robust pedestrian attribute recognition via fixed orthogonal classifier.

Abstract

Pedestrian attribute recognition (PAR) aims to achieve robust prediction of multiple binary attributes in pedestrian images. Traditional PAR methods employ joint learning of a backbone network to extract shared features and a linear classifier to predict binary attributes. However, this paradigm faces key challenges: (1) the shared feature category grows exponentially with the number of attribute combinations up to 2^C categories where C is the total number of attributes, (2) significant intra-class variance leads to uneven feature distributions, and (3) statistical correlations among attributes cause classifier confusion, hindering generalization. To address these issues, we propose a novel Attribute-Specialized Representation Learning (ASRL) framework. Our framework employs an efficient split-concat-project module with a fixed orthogonal classifier to focus on attribute-specific traits, thereby reducing classifier confusion caused by statistical correlations among attributes. Additionally, we incorporate two regularization terms to minimize intra-class variance and align attribute-specialized features with the fixed orthogonal classifier, ensuring structural separation of attribute-specialized features. Extensive experiments on multiple benchmark datasets demonstrate that our method outperforms state-of-the-art methods. Furthermore, our method achieves significant improvements on the cross-domain UPAR* dataset, showcasing its robustness and generalizability.