Heterogeneous Feature Re-Sampling for Balanced Pedestrian Attribute Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Pedestrian attribute recognition (PAR) encompasses diverse descriptors, from biometrics to accessories.
• Real-world datasets often feature under-represented attributes, leading to significant data imbalance.
• Existing PAR approaches can be over-specialized, failing to address the fundamental challenge of imbalanced data.

Purpose of the Study:

• To reframe PAR as a multi-label recognition task under severe data imbalance.
• To develop novel methods that achieve label-balanced learning independent of attribute co-occurrence.
• To enhance the robustness and performance of PAR systems on diverse, realistic datasets.

Main Methods:

• Feature re-sampled detached learning (FRDL) is proposed to balance attribute sampling distributions without affecting co-occurring label priors.
• Gradient-oriented augment translating (GOAT) is introduced to mitigate feature noise and semantic imbalance exacerbated by FRDL.
• FRDL and GOAT are integrated into a unified framework for comprehensive performance improvement.

Main Results:

• The integrated FRDL and GOAT framework significantly advances the state-of-the-art in PAR.
• Performance improvements are demonstrated across various realistic benchmarks.
• The methods achieve these gains with a minimal increase in computational cost.

Conclusions:

• This work establishes the first label-independent and impartial balanced learning approach for PAR.
• The proposed methods effectively address the challenges of data imbalance and attribute co-occurrence in PAR.
• The advancements offer a more fundamental and robust perspective on pedestrian attribute recognition.