Adapting Frozen Large Vision Models for Silhouette-Based Gait Recognition

Abstract

In practical gait recognition scenarios where only silhouette sequences are available without original videos, directly applying appearance-driven BigGait frameworks often leads to a mismatch between methodology and data conditions. To leverage the prior advantages of frozen Large Vision Models (LVMs) for universal feature extraction, this research proposes a silhouette-adapted BigGait scheme. Specifically, single-channel silhouettes are first mapped into a three-channel space to match the DINOv2 pre-training distribution, with a deterministic foreground prior constructed to guide token selection. Subsequently, a Part-aware Temporal Aggregation (PTA) mechanism is introduced in the downstream encoding stage to strengthen fine-grained dynamic representation through adaptive weighting. To enhance robustness against contour noise, a dual-view silhouette perturbation task is constructed, with dual-level consistency constraints applied across both global and part-level representations. Experiments on the CASIA-B benchmark demonstrate that the proposed scheme improves recognition performance under controlled walking conditions, especially under clothing changes. These results provide preliminary evidence that frozen vision priors can be adapted to silhouette-only gait recognition when the modality gap is explicitly addressed.