Towards trustworthy gait recognition: visualizing and refining hierarchical motion dependencies

Deep learning-based gait recognition has achieved strong identification accuracy, yet its deployment in high-stakes forensic and clinical scenarios is hindered by limited auditability and sensitivity to covariate disturbances (e.g., clothing occlusion and carried objects). This paper proposes the Interpretable Hierarchical Gait Refinement Network (IHGR-Net), an end-to-end hierarchical silhouette-based framework that makes both spatial cue reliability (“where”) and temporal cue selection (“when”) explicit and uses them to refine representation learning. Specifically, a Saliency-Guided Reliability Assessment (SGRA) module converts part-level activation/saliency responses into reliability scores and performs reliability-gated refinement over anatomically aligned partitions, down-weighting corrupted regions while emphasizing stable cues. In addition, a Semantic Key-Frame Aligner (SKA) replaces opaque temporal pooling with an attention scorer that yields an interpretable frame-importance distribution and selects biomechanically informative frames. Experiments on CASIA-B and OU-MVLP show that IHGR-Net improves robustness to clothing and carrying variations while maintaining strong clean-condition performance, achieving 98.3/96.5/89.9% Rank-1 on CASIA-B (NM/BG/CL) and 90.3% on OU-MVLP.