RobustGait: Robustness Analysis for Appearance Based Gait Recognition

TL;DR

RobustGait provides a comprehensive framework to assess appearance-based gait recognition robustness under real-world degradations by injecting RGB-level noise and tracing its propagation through silhouette extraction to final recognition. It benchmarks 15 corruption types at five severities across CASIA-B, CCPG, SUSTech1K with MEVID in-the-wild validation, evaluating six gait models and several silhouette extractors under cross-extractor and deployment scenarios. Key findings show silhouette extractor biases can dominate performance, robustness depends on perturbation type and architecture, and strategies like noise-aware training and LoRA-based distillation can enhance resilience while preserving clean accuracy. The work offers a practical, scalable protocol for robust gait benchmarking with actionable guidance for deploying appearance-based gait systems in real-world settings.

Abstract

Appearance-based gait recognition have achieved strong performance on controlled datasets, yet systematic evaluation of its robustness to real-world corruptions and silhouette variability remains lacking. We present RobustGait, a framework for fine-grained robustness evaluation of appearance-based gait recognition systems. RobustGait evaluation spans four dimensions: the type of perturbation (digital, environmental, temporal, occlusion), the silhouette extraction method (segmentation and parsing networks), the architectural capacities of gait recognition models, and various deployment scenarios. The benchmark introduces 15 corruption types at 5 severity levels across CASIA-B, CCPG, and SUSTech1K, with in-the-wild validation on MEVID, and evaluates six state-of-the-art gait systems. We came across several exciting insights. First, applying noise at the RGB level better reflects real-world degradation, and reveal how distortions propagate through silhouette extraction to the downstream gait recognition systems. Second, gait accuracy is highly sensitive to silhouette extractor biases, revealing an overlooked source of benchmark bias. Third, robustness is dependent on both the type of perturbation and the architectural design. Finally, we explore robustness-enhancing strategies, showing that noise-aware training and knowledge distillation improve performance and move toward deployment-ready systems.

Figures (12)

• Figure 1: Overlooked biases in appearance-based gait benchmarks. (i) Variation across silhouette extractors leads to evaluation bias due to variable silhouette quality, motivating the need for standardized extraction across benchmarks. (ii) directly applying noise to silhouettes restricts corruptions to simple augmentations as flipping, rotation, or erasing, whereas injecting noise at the RGB level allows various temporal, environmental, and digital degradations to propagate to silhouettes, better reflecting real-world scenarios.
• Figure 2: Overview of noises: Qualitative visualization of four major taxonomy of noises studied in our benchmark.
• Figure 3: Qualitative analysis of parsing models on CASIA-B. Silhouette quality decreases from left to right. M2FP, SCHP, and GSAM preserve body structure, while CDGNet and STVC show degradation.
• Figure 4: Impact of silhouette segmentation on gait recognition. Left (CASIA-B and Right (CCPG). The IoU curve positively correlates with recognition performance: segmentation methods with higher IoU generally yield higher Rank-1 accuracy across gait models (e.g., high-IoU M2FP on CASIA-B; SCHP on CCPG), while lower-IoU methods (e.g., CDGNet) correspond to lower accuracies. This highlights that better silhouette masks improve downstream gait recognition on those silhouettes.
• Figure 5: Impact of silhouette segmentation on gait recognition on SUSTECH.Legend identical to Fig. \ref{['fig:segmentation1']}