Progressive Feature Learning for Realistic Cloth-Changing Gait Recognition

Existing datasets and methods for gait recognition struggle to address the challenging issue of cloth-changing scenarios effectively. In practice, gait recognition models are often trained using automatically labeled data, where sequences are constrained by specific view and clothing conditions. Specifically, cross-view datasets typically include sequences under normal walking conditions without cloth changes. In contrast, cross-cloth datasets feature clothing variations but are limited to front views. This imbalance leads to suboptimal performance under realistic conditions. To tackle this issue, we define the task as Realistic Cloth-Changing Gait Recognition (abbreviated as RCC-GR) and we construct two benchmarks: CASIA-BN-RCC and OUMVLP-RCC, to simulate real-world scenarios. Furthermore, we propose a novel framework, Progressive Feature Learning, compatible with existing backbone models, to improve recognition performance in RCC-GR. Our framework includes two key components: Progressive Mapping and Progressive Uncertainty. These techniques first focus on extracting cross-view features and then incorporate cross-cloth features, ensuring that features from the cross-view dataset dominate the feature space. This approach mitigates the uneven distribution caused by cross-cloth variations. Experiments conducted on our benchmarks demonstrate that the proposed framework significantly improves recognition accuracy, particularly under cloth-changing conditions.