High-order diversity feature learning for pedestrian attribute recognition

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Pedestrian attribute recognition (PAR) is crucial for understanding urban environments.
• Existing part-based and attention-based methods for PAR have limitations in detail capture and global context integration.
• Current methods often struggle with fine-grained details and can be affected by irrelevant information.

Purpose of the Study:

• To develop a novel method for pedestrian attribute recognition (PAR) that overcomes limitations of existing approaches.
• To enhance the extraction of fine-grained, attribute-specific features by integrating high-order statistics and global context.
• To improve the accuracy and robustness of PAR systems.

Main Methods:

• Proposed a High-order Diversity Feature Learning (HDFL) method for PAR, leveraging Vision Transformer (ViT).
• Introduced an Attribute-specific Detailed Feature Exploration (ADFE) module using a polynomial predictor to capture high-order statistics and fine-grained features.
• Developed a Soft-redundancy Perception Loss (SPLoss) to promote feature diversity by measuring feature redundancy.

Main Results:

• The proposed HDFL method achieved state-of-the-art (SOTA) performance on multiple PAR datasets.
• Outperformed previous SOTA on the challenging PA100K dataset by 1.69%, achieving a mean accuracy (mA) of 84.92%.
• Demonstrated the effectiveness of the ADFE module in generating detailed, attribute-specific features.

Conclusions:

• The HDFL method offers a significant advancement in pedestrian attribute recognition.
• Integrating detailed attention with global context effectively addresses limitations of prior PAR techniques.
• The proposed approach provides a robust and accurate solution for complex PAR tasks.