DAGait: Generalized Skeleton-Guided Data Alignment for Gait Recognition

ICME 2025
Zhengxian Wu1*,  Chuanrui Zhang1*,  Hangrui Xu2,  Peng Jiao1,  Haoqian Wang1 
1Tsinghua University,  2Hefei University of Technology 
* Equal Contribution 
Paper Code
architecture
Overview of DAGait. In the data preprocessing stage, silhouettes and skeletons are extracted from the RGB image sequences. During the data alignment phase, prior spatial relationships between skeleton joints and silhouette regions are used to apply an affine transformation, correcting the silhouette. Finally, the aligned silhouette images are input into the backbone network for recognition.

TL;DR

We present DAGait, a universal data alignment strategy for gait recognition, to alleviate spatiotemporal distribution inconsistencies.

Abstract

Gait recognition is emerging as a promising and innovative area within the field of computer vision, widely applied to remote person identification. Although existing gait recognition methods have achieved substantial success in controlled laboratory datasets, their performance often declines significantly when transitioning to wild datasets. We argue that the performance gap can be primarily attributed to the spatio-temporal distribution inconsistencies present in wild datasets, where subjects appear at varying angles, positions, and distances across the frames. To achieve accurate gait recognition in the wild, we propose a skeleton-guided silhouette alignment strategy, which uses prior knowledge of the skeletons to perform affine transformations on the corresponding silhouettes. To the best of our knowledge, this is the first study to explore the impact of data alignment on gait recognition. We conducted extensive experiments across multiple datasets and network architectures, and the results demonstrate the significant advantages of our proposed alignment strategy. Specifically, on the challenging Gait3D dataset, our method achieved an average performance improvement of 7.9\% across all evaluated networks. Furthermore, our method achieves substantial improvements on cross-domain datasets, with accuracy improvements of up to 24.0\%.

Introduction

Intro
we propose a gait recognition framework, named DAGait, designed to address the distribution discrepancies commonly observed in wild datasets. Accurate data alignment can mitigate interference from perspective shifts and posture variations, enabling the network to learn universal gait features. (a) The figure demonstrates the silhouette and gait energy image (GEI) within a sequence before and after data alignment, highlighting the alignment's effectiveness in mitigating spatio-temporal distribution inconsistencies. (b) The performance comparison of GaitBase without and with data alignment across various datasets, demonstrating significant accuracy improvements, particularly on the Gait3D wild dataset.

Comparisons with the State-of-the-art

The proposed data alignment strategy consistently improves the performance of existing methods across all three datasets, with particularly significant effects observed on the more challenging wild dataset, Gait3D.

SOTA comparisons

Visualization of Alignment

We present the visualizations of the aligned silhouettes and GEIs: