Learning Gait Representation From Massive Unlabelled Walking Videos: A Benchmark

Area of Science:

• Biometrics and Pattern Recognition
• Computer Vision
• Machine Learning

Background:

• Gait recognition, a promising biometric, faces challenges due to sensitivity to variations and high annotation costs.
• Existing methods often require extensive labeled data, limiting practical applications.
• Developing methods to learn robust gait representations from unlabelled data is crucial.

Purpose of the Study:

• To propose a large-scale self-supervised benchmark for gait recognition using contrastive learning.
• To learn general gait representations from massive unlabelled walking videos.
• To reduce the dependency on costly annotated data for gait identification.

Main Methods:

• Collected GaitLU-1M, a large-scale unlabelled dataset with 1.02 million walking sequences.
• Developed GaitSSB, a self-supervised baseline model utilizing contrastive learning.
• Evaluated the pre-trained model on four benchmark datasets (CASIA-B, OU-MVLP, GREW, Gait3D).

Main Results:

• Unsupervised results from GaitSSB are comparable or superior to early methods.
• With transfer learning, GaitSSB significantly outperforms existing methods on benchmarks.
• Pre-training demonstrated substantial savings in annotation costs (50-80%) for downstream tasks.

Conclusions:

• GaitLU-1M is the first large-scale unlabelled gait dataset, and GaitSSB is a pioneering method for unsupervised gait recognition.
• The proposed self-supervised approach effectively learns general gait representations, enhancing identification performance.
• This work offers insights into gait-specific contrastive learning and its potential for practical biometric systems.