Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Silhouette-based Gait Foundation Model

Dingqiang Ye, Chao Fan, Kartik Narayan, Bingzhe Wu, Chengwen Luo, Jianqiang Li, Vishal M. Patel

TL;DR

FoundationGait presents a scalable, self-supervised gait foundation model trained on WebGait-2M to unify recognition and healthcare tasks. By preserving part-level diversity through a part-aware training strategy and a teacher-student EMA framework, it demonstrates scaling laws in gait modeling and robust cross-task generalization, achieving leading zero-shot results on Gait3D and OU-MVLP, as well as strong improvements in healthcare-related gait analyses. The work provides extensive empirical validation, including ablations, and highlights both the practical potential and remaining challenges, such as exploring ViT-based gait foundations and scaling beyond 1B parameters. Overall, FoundationGait advances toward a unified gait foundation model with broad applicability and practical impact for large-scale gait analytics.

Abstract

Gait patterns play a critical role in human identification and healthcare analytics, yet current progress remains constrained by small, narrowly designed models that fail to scale or generalize. Building a unified gait foundation model requires addressing two longstanding barriers: (a) Scalability. Why have gait models historically failed to follow scaling laws? (b) Generalization. Can one model serve the diverse gait tasks that have traditionally been studied in isolation? We introduce FoundationGait, the first scalable, self-supervised pretraining framework for gait understanding. Its largest version has nearly 0.13 billion parameters and is pretrained on 12 public gait datasets comprising over 2 million walking sequences. Extensive experiments demonstrate that FoundationGait, with or without fine-tuning, performs robustly across a wide spectrum of gait datasets, conditions, tasks (e.g., human identification, scoliosis screening, depression prediction, and attribute estimation), and even input modality. Notably, it achieves 48.0% zero-shot rank-1 accuracy on the challenging in-the-wild Gait3D dataset (1,000 test subjects) and 64.5% on the largest in-the-lab OU-MVLP dataset (5,000+ test subjects), setting a new milestone in robust gait recognition. Coming code and model: https://github.com/ShiqiYu/OpenGait.

Silhouette-based Gait Foundation Model

TL;DR

FoundationGait presents a scalable, self-supervised gait foundation model trained on WebGait-2M to unify recognition and healthcare tasks. By preserving part-level diversity through a part-aware training strategy and a teacher-student EMA framework, it demonstrates scaling laws in gait modeling and robust cross-task generalization, achieving leading zero-shot results on Gait3D and OU-MVLP, as well as strong improvements in healthcare-related gait analyses. The work provides extensive empirical validation, including ablations, and highlights both the practical potential and remaining challenges, such as exploring ViT-based gait foundations and scaling beyond 1B parameters. Overall, FoundationGait advances toward a unified gait foundation model with broad applicability and practical impact for large-scale gait analytics.

Abstract

Gait patterns play a critical role in human identification and healthcare analytics, yet current progress remains constrained by small, narrowly designed models that fail to scale or generalize. Building a unified gait foundation model requires addressing two longstanding barriers: (a) Scalability. Why have gait models historically failed to follow scaling laws? (b) Generalization. Can one model serve the diverse gait tasks that have traditionally been studied in isolation? We introduce FoundationGait, the first scalable, self-supervised pretraining framework for gait understanding. Its largest version has nearly 0.13 billion parameters and is pretrained on 12 public gait datasets comprising over 2 million walking sequences. Extensive experiments demonstrate that FoundationGait, with or without fine-tuning, performs robustly across a wide spectrum of gait datasets, conditions, tasks (e.g., human identification, scoliosis screening, depression prediction, and attribute estimation), and even input modality. Notably, it achieves 48.0% zero-shot rank-1 accuracy on the challenging in-the-wild Gait3D dataset (1,000 test subjects) and 64.5% on the largest in-the-lab OU-MVLP dataset (5,000+ test subjects), setting a new milestone in robust gait recognition. Coming code and model: https://github.com/ShiqiYu/OpenGait.

Paper Structure

This paper contains 17 sections, 5 equations, 6 figures, 16 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Works
  3. Method
  4. Scaling Issue in Gait Models
  5. FoundationGait
  6. Experiments
  7. Datasets
  8. Implementation Details
  9. Zero-shot Performance
  10. Fine-tuning on Gait Recognition
  11. Fine-tuning on Healthcare Tasks
  12. Ablation Study
  13. Conclusion and Limitations
  14. Supplementart Material
  15. Downstream Fine-tuning
  16. ...and 2 more sections

Figures (6)

  • Figure 1: FoundationGait: A Scalable and Unified Gait Foundation Model.
  • Figure 2: Unscalable Issue on the Self-supervised GaitSSB fan2023learning. All models are pretrained on WebGait-2M.
  • Figure 3: Unscalable Issue on the Supervised DeepGaitV2 fan2025opengait.
  • Figure 4: Overview of FoundationGait. A scalable self-supervised gait learning framework.
  • Figure 5: Recognition Head during Fine-tuning.
  • ...and 1 more figures