RestoreGait: a recognition-oriented temporal inpainting framework for occluded gait recognition

Abstract

Gait recognition represents a nonintrusive biometric modality that suffers from substantial performance degradation when subjected to real-world occlusions, such as those caused by bushes, pillars, or crowds. Current approaches either passively suppress occluded regions or employ image inpainting techniques that fail to account for temporal coherence and identity consistency. We introduce RestoreGait, an end-to-end framework designed to actively recover occluded gait cues through a lightweight pseudo-3D spatiotemporal decoupled inpainting module. This module decouples spatial contour restoration from temporal motion aggregation, thereby effectively utilizing visible information across multiple frames. Furthermore, an uncertainty-aware soft-gating mechanism is employed to suppress inpainting artifacts of low confidence, whereas a semantic identity consistency loss ensures that the restored sequences maintain discriminative identity features within the deep embedding space, rather than merely achieving visual realism. Extensive experiments on the SUSTech1K, OccGait, Gait3D, and GREW datasets demonstrate that the proposed method yields consistent improvements over strong baselines with negligible computational overhead, demonstrating strong and consistent performance on occluded gait benchmarks.