On Learning Disentangled Representations for Gait Recognition.

Area of Science:

• Computer Vision
• Biometrics
• Machine Learning

Background:

• Gait recognition is a key biometric modality, but current methods struggle with variations like clothing and viewing angles.
• Existing approaches using silhouettes or body models show degraded performance under confounding variables.
• Frontal-view gait recognition is particularly challenging due to limited available gait cues.

Purpose of the Study:

• To propose a novel AutoEncoder framework, GaitNet, for robust gait recognition.
• To explicitly disentangle appearance, canonical, and pose features from RGB imagery.
• To introduce and utilize a new Frontal-View Gait (FVG) dataset for evaluating frontal gait recognition.

Main Methods:

• Developed GaitNet, an AutoEncoder framework to disentangle gait features.
• Integrated Long Short-Term Memory (LSTM) for dynamic pose feature integration.
• Averaged canonical features for static gait representation.
• Collected and utilized the Frontal-View Gait (FVG) dataset.

Main Results:

• GaitNet demonstrated superior performance compared to state-of-the-art (SOTA) methods on CASIA-B, USF, and FVG datasets.
• Qualitative analysis confirmed the framework's ability for effective feature disentanglement.
• The method showed promising computational efficiency and advantages over face recognition in specific scenarios.

Conclusions:

• GaitNet offers a robust solution for gait recognition, effectively handling variations and challenging views.
• The proposed feature disentanglement approach significantly improves recognition accuracy.
• Gait biometrics show potential advantages over other methods in scenarios with long-distance or low-resolution imagery.