GaitSnippet: Gait Recognition Beyond Unordered Sets and Ordered Sequences

Saihui Hou, Chenye Wang, Wenpeng Lang, Zhengxiang Lan, Yongzhen Huang^†

[📜Paper] [📝Arxiv]
We propose GaitSnippet, a novel snippet-based paradigm for gait recognition, treating a gait sequence as a union of individualized actions to capture multi-scale temporal contexts. Based on the new modeling concept called snippet, we design speical sampling and modeling methods, through extensive experiments shows SOTA performance using only with 2D CNN backbone.

Gait Transformer: End-to-End Transformer Backbone for Gait Recognition

Saihui Hou, Wenpeng Lang, Jilong Wang, Yan Huang, Liang Wang, Yongzhen Huang^†

[📜Paper] [🖼️Poster]
We propose GaT, an end-to-end Transformer for gait recognition. It addresses spatial-temporal dynamics, fine-grained motion, and data scarcity via three modules: hybrid patch embedding with group-batch bormalization, decomposed token mixer for context dependencies, and hybrid positional encoding strategy. Without any pretraining, GaT achieves SOTA performance on popular datasets.

Beyond Sparse Keypoints: Dense Pose Modeling for Robust Gait Recognition

Wenpeng Lang, Saihui Hou, Yongzhen Huang^†

[📜Paper] [🌐Website] [🎬Video] [🧩Code]
We propose DPGait, a dense pose-based method to solve the limitations of sparse keypoints. On the upstream, we extend estimation model to output human dense points. On the downstream, we design a divide-and-conquer modeling architecture. Our method achieves SOTA performance across three datasets, demonstrating the effectiveness of method in complex scenarios.