Exploring Self-Supervised Vision Transformers for Gait Recognition in the Wild.

Area of Science:

• Computer Vision
• Biometrics
• Machine Learning

Background:

• Gait analysis offers unobtrusive, remote biometric identification without subject cooperation.
• Traditional gait recognition relies on controlled data, limiting real-world applicability.
• Self-supervised learning on large datasets enhances gait representation robustness.

Purpose of the Study:

• To investigate the efficacy of various vision transformer architectures for self-supervised gait recognition.
• To evaluate transformer performance on diverse, large-scale gait datasets.
• To analyze the impact of spatial and temporal information on transformer-based gait recognition.

Main Methods:

• Adaptation and pretraining of five vision transformer models (ViT, CaiT, CrossFormer, Token2Token, TwinsSVT) using self-supervised learning.
• Utilizing large-scale gait datasets GREW and DenseGait for pretraining.
• Extensive evaluation on benchmark datasets CASIA-B and FVG using zero-shot and fine-tuning approaches.

Main Results:

• Vision transformers show promise for self-supervised gait recognition.
• Hierarchical transformer models (e.g., CrossFormer) demonstrate superior performance in processing finer-grained motion compared to whole-skeleton approaches.
• Analysis reveals the relationship between spatial/temporal information utilization and transformer effectiveness.

Conclusions:

• Transformer architectures are viable for self-supervised gait recognition.
• Hierarchical designs are beneficial for capturing detailed motion dynamics in gait analysis.
• Self-supervised learning with transformers offers a robust pathway for advanced biometric gait recognition.