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

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Pedestrian attribute recognition (PAR) traditionally uses joint learning, facing challenges with feature category explosion, intra-class variance, and attribute correlations.
• Existing methods struggle with exponential category growth (2^C) and classifier confusion due to attribute statistical dependencies.

Purpose of the Study:

• To propose a novel Attribute-Specialized Representation Learning (ASRL) framework to overcome limitations in traditional PAR methods.
• To enhance the robustness and generalizability of pedestrian attribute recognition.

Main Methods:

• Developed an Attribute-Specialized Representation Learning (ASRL) framework utilizing a split-concat-project module and a fixed orthogonal classifier.
• Incorporated regularization terms to minimize intra-class variance and align attribute-specialized features, ensuring structural separation.

Main Results:

• The proposed ASRL framework significantly outperforms state-of-the-art methods on multiple benchmark datasets.
• Demonstrated substantial improvements on the cross-domain UPAR* dataset, highlighting robustness and generalizability.

Conclusions:

• ASRL effectively addresses challenges in PAR by focusing on attribute-specific traits and reducing classifier confusion.
• The framework offers a more robust and generalizable solution for pedestrian attribute recognition tasks.