OpenGait: Revisiting Gait Recognition Toward Better Practicality

Chao Fan, Junhao Liang, Chuanfu Shen, Saihui Hou, Yongzhen Huang, Shiqi Yu

TL;DR

Gait recognition models suffer a massive ~40% accuracy drop when moving from controlled laboratory environments to the real world. This paper introduces OpenGait, a unified benchmarking framework that reveals many complex, state-of-the-art modules actually hinder real-world performance. Leveraging these insights, the authors propose GaitBase—a strikingly simple baseline model that strips away unnecessary architectural "bells and whistles" to set a new standard for practical, outdoor gait recognition.

What Problem Does This Paper Address?

Gait recognition—identifying individuals by the unique way they walk—is a powerful biometric technology because it works from a distance without requiring physical contact or high-resolution facial features. While deep learning models have achieved near-perfect accuracy on indoor laboratory datasets, their performance plummets when deployed "in the wild."

The authors identify a critical disconnect in current research: most existing models are over-engineered to exploit the clean, artifact-free silhouettes of indoor datasets. When faced with real-world challenges like complex backgrounds, lighting changes, and occlusions, these specialized architectures fail. Furthermore, the community has lacked a unified, open-source codebase to fairly evaluate different methods across both indoor and outdoor scenarios.

Figure 1. Performance of popular models and ours baseline on 4 major gait datasets [2, 6, 12, 13, 19, 45]. The left two are indoor datasets [34,41], the right two are outdoor datasets [45,48].

Methods

To address these issues, the authors first built OpenGait, a flexible and efficient PyTorch-based codebase designed to support multiple input modalities (silhouettes, 2D/3D skeletons, 3D meshes) and various learning paradigms.

Using OpenGait to systematically strip down and re-evaluate previous state-of-the-art (SOTA) models, they developed a streamlined baseline called GaitBase. Unlike its predecessors, GaitBase avoids intricate, custom-designed modules. Its pipeline consists of:

  1. Strong Spatial Data Augmentation: Incorporating horizontal flipping, rotation, and perspective transformations to simulate real-world viewpoint and camera height variations.
  2. A Robust Backbone: Replacing shallow, custom convolutional networks with a standard, deeper ResNet9 architecture.
  3. Simple Temporal and Horizontal Pooling: Aggregating frame-level features across time and slicing them horizontally to capture basic body-part relations, completely omitting complex multi-scale pyramid structures.
  4. Standard Loss Functions: Training via a combination of standard triplet loss and cross-entropy loss adjusted by a BNNeck layer.
Figure 2. The design principles of proposed codebase OpenGait. Seg is for segmentation, Rec is for recognition, and Br is for branch.

Datasets

The paper evaluates models across four major public datasets, highlighting the stark contrast between indoor and outdoor environments:

  • Indoor (Controlled):
  • CASIA-B: A widely used but small dataset (124 subjects) featuring clean silhouettes generated by outdated background subtraction.
  • OU-MVLP: A massive indoor dataset (over 10,000 subjects) but limited to a single walking status without clothing or carrying variations.
  • Outdoor (In the Wild):
  • GREW: The largest outdoor dataset, containing over 26,000 sequences from 882 cameras with diverse real-world attributes (carrying, dressing, age groups).
  • Gait3D: A large-scale outdoor dataset captured in a supermarket, providing challenging low-quality segmentations and 3D mesh data.

Key Insights and Ideas

  • The Complexity Trap: Through extensive ablation studies, the authors discovered that highly praised modules from prior SOTA models—such as GaitPart's Focal Convolutions or GaitSet's Global Pipeline—actually degrade performance on outdoor datasets. These modules rely on pristine silhouettes; when segmentations are noisy or imperfect, direct feature-splitting introduces errors.
  • Backbones Over "Bells and Whistles": Previous research over-estimated the value of custom modules because they evaluated them on weak, shallow backbones. Upgrading to a robust, standard backbone like ResNet9 naturally absorbs the challenges of noisy data, rendering complex custom modules redundant.
  • The Indoor Benchmark Bias: Strategies that optimize outdoor performance (like aggressive data augmentation and varying training sequence lengths) can slightly decrease accuracy on indoor datasets. This reveals that optimizing purely for indoor benchmarks leads to severe real-world over-fitting.

Key Results

GaitBase demonstrates that simplicity can outperform complexity in practical scenarios. While maintaining highly competitive results on indoor datasets, it achieves a massive performance leap on outdoor data:

  • GREW Dataset: GaitBase achieves a Rank-1 accuracy of 60.1%, outperforming the best previous model (CSTL at 50.6%) by +9.5%.
  • Gait3D Dataset: GaitBase reaches 64.6% Rank-1 accuracy, trouncing the previous SOTA (SMPLGait at 46.3%) by a staggering +18.3%.
Figure 4. The pipeline of GaitBase. DA is for data augmentation, ResNet9 is for ResNet-like backbone, TP is for temporal pooling, HP is for horizontal pooling (w/o pyramid structure [9]).

Limitations and Future Work

Despite setting a new state-of-the-art, a 60% accuracy rate highlights that practical gait recognition is far from solved. The authors note several open avenues for the community:

  1. Gait Verification: Most research focuses on identification (retrieval), ignoring the binary verification scenario (matching two gaits), which is highly impacted by intra-class variations like changing clothes.
  2. Architecture Shift: Future work should explore transitioning the gait recognition backbone from CNNs to Transformers to better capture long-range spatial-temporal dependencies.
  3. Unsupervised Learning: Collecting annotated data in the wild is expensive. Developing unsupervised methods to learn general gait representations from massive, unlabelled video streams is a crucial next step.

Summary updated

Abstract

Gait recognition is one of the most critical long-distance identification technologies and increasingly gains popularity in both research and industry communities. Despite the significant progress made in indoor datasets, much evidence shows that gait recognition techniques perform poorly in the wild. More importantly, we also find that some conclusions drawn from indoor datasets cannot be generalized to real applications. Therefore, the primary goal of this paper is to present a comprehensive benchmark study for better practicality rather than only a particular model for better performance. To this end, we first develop a flexible and efficient gait recognition codebase named OpenGait. Based on OpenGait, we deeply revisit the recent development of gait recognition by re-conducting the ablative experiments. Encouragingly,we detect some unperfect parts of certain prior woks, as well as new insights. Inspired by these discoveries, we develop a structurally simple, empirically powerful, and practically robust baseline model, GaitBase. Experimentally, we comprehensively compare GaitBase with many current gait recognition methods on multiple public datasets, and the results reflect that GaitBase achieves significantly strong performance in most cases regardless of indoor or outdoor situations. Code is available at https://github.com/ShiqiYu/OpenGait.
BibTeX
@inproceedings{fan2022opengait,
  author = {Chao Fan and Junhao Liang and Chuanfu Shen and Saihui Hou and Yongzhen Huang and Shiqi Yu},
  title = {{OpenGait: Revisiting Gait Recognition Toward Better Practicality}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2022},
  url = {https://openreview.net/forum?id=TVOu4bjH7W_},
}
Figure