Source-linked AI summary
OpenGait: Revisiting Gait Recognition Toward Better Practicality
Chao Fan, Junhao Liang, Chuanfu Shen, Saihui Hou, Yongzhen Huang, Shiqi Yu
TL;DR
Indoor progress in gait recognition does not fully generalize to outdoor datasets, and prior ablative conclusions can vary across datasets. The paper develops OpenGait to benchmark methods fairly, re-evaluates prior findings, and introduces GaitBase, which performs strongly across indoor and outdoor datasets. The authors nevertheless report a significant gap before gait recognition is accurate enough for real-world applications.
Problem
Indoor gait-recognition findings may not generalize to outdoor settings, while the field lacks a unified and fair platform for practical evaluation.
Method
The paper develops OpenGait, reimplements and re-evaluates prior methods and ablations, and uses the findings to develop the simple GaitBase baseline.
Results
GaitBase achieves the best performance on most gait datasets, especially on datasets in the wild.
Takeaways & Limitations
OpenGait provides a fair research platform, while GaitBase can serve as a new baseline for future gait-recognition studies.
Takeaways & Limitations
GREW rank-1 accuracy reaches 60.1%, but a significant gap remains before gait recognition is accurate enough for real-world applications.
Abstract
from arXiv · showhide
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.
1. Introduction
The paper argues that gait recognition advances on indoor datasets do not reliably transfer to outdoor settings or real applications. It responds with a practical benchmark, OpenGait, re-evaluation of prior findings, and the GaitBase baseline.
- Motivation: Gait recognition can identify people from a distance in uncontrolled, contactless settings using walking characteristics.The paper highlights applications including public security, criminal investigation, and suspect tracking.
- Practicality Gap: Over 40% accuracy degradation occurs for most existing gait recognition methods when moving from indoor to outdoor datasets.The paper attributes the gap mainly to occlusion, background variation, and illumination changes.
- Benchmark Motivation: Ablative conclusions from prior works vary across datasets, motivating a comprehensive benchmark study focused on practical gait recognition.The study specifically re-examines findings on recently built outdoor datasets.
- Contributions: OpenGait is a flexible PyTorch codebase designed to support diverse datasets, state-of-the-art methods, and multiple learning frameworks.Supported examples include indoor and outdoor datasets, end-to-end, multimodal, and contrastive-learning paradigms.
- Re-evaluation: Reimplemented ablations show that several branches reported as beneficial indoors do not exhibit superiority on outdoor datasets.The paper names GaitSet's MGP, GaitPart's FConv, GaitGL's local-feature branch, and SMPLGait's SMPL branch.
- Baseline Model: GaitBase is a simple baseline that achieves comparable or superior indoor performance and outperforms recent methods on wild datasets.The paper presents it as structurally simple, experimentally powerful, and empirically robust.
2. Related Work
Related work spans model-based and appearance-based gait recognition, alongside diverse architectures, modalities, and training frameworks. OpenGait is positioned as a compatible platform for reanalyzing these methods across indoor and outdoor settings.
- Method Taxonomy: Model-based methods use estimated human structure, such as 2D/3D pose or SMPL, while appearance-based methods learn from visual gait representations.Examples include PoseGait and GaitGraph among model-based approaches.
- Appearance-Based Methods: Prior appearance-based methods introduce global-local convolution, multiscale temporal features, or adaptive 3D local operations to model gait information.The cited examples are GaitGL, CSTL, and 3DLocal.
- Other Frameworks: Other approaches use edge-trainable intermediate modalities or contrastive learning over millions of unlabeled gait sequences.GaitEdge and GaitSSB illustrate these directions.
- OpenGait Context: OpenGait is compatible with nearly all of these methods and supports experimental reanalysis as research shifts from indoor to outdoor environments.The platform is presented as infrastructure for systematic comparison and new insights.
3. OpenGait
OpenGait addresses the lack of a unified, fair evaluation platform by integrating diverse gait modalities, frameworks, datasets, methods, and evaluation protocols. Its modular PyTorch design separates data, modeling, and evaluation responsibilities.
- Purpose: OpenGait provides a unified and fair platform intended to support both academic research and practical gait-recognition applications.The codebase is presented as a dependable solution to fragmented evaluation infrastructure.
- Compatibility: OpenGait supports silhouette, 2D/3D skeleton, SMPL-parameter, and RGB input modalities.This extends beyond repositories that typically support only one modality.
- Compatibility: OpenGait supports multimodal, end-to-end, and contrastive-learning frameworks, enabling broader method integration.These frameworks are identified as increasingly common directions in gait recognition.
- Datasets: The toolbox fully supports indoor CASIA-B and OU-MVLP and outdoor GREW and Gait3D datasets, including preprocessing through evaluation.It is also compatible with the HID and GREW competitions.
- Methods: OpenGait reproduces numerous state-of-the-art methods, including GaitSet, GaitPart, GaitGL, Gait3D, SMPLGait, GaitEdge, and GaitSSB.This supports fair comparison within one evaluation environment.
- Architecture: The codebase is organized into data, modeling, and evaluation modules.The data module handles loading, sampling, and preprocessing; modeling exposes customizable backbone, neck, head, and loss components; evaluation unifies dataset protocols.
4. Revisit Deep Gait Recognition
OpenGait enables fair outdoor re-evaluation of established gait-recognition components, revealing that several indoor-derived design choices do not improve performance on practical gait data. The analysis motivates stronger backbones, broader evaluation, and simpler architectures.
- Experimental Recheck on Previous Methods: OpenGait re-conducts ablation studies on Gait3D to test whether conclusions established mainly on indoor datasets remain robust in practical settings.The study contrasts results on outdoor Gait3D with prior evaluations centered on CASIA-B and OU-MVLP.
- GaitSet: Removing MGP or HPP’s multi-scale mechanism matches or improves GaitSet performance on CASIA-B and Gait3D while saving over 80% of training weights.The authors attribute the result partly to weak statistical pooling functions that fail to learn additional knowledge from multi-scale representations.
- GaitPart: Rank-1: +7.0% on Gait3D when GaitPart’s FConv is replaced with regular convolution, indicating that feature-map splitting can hurt wild-data recognition.The paper links this effect to low-quality segmentation in outdoor data and notes that sequential inputs may also be vulnerable to frame drops and walking-speed changes.
- GaitGL: Removing GaitGL’s local branch improves performance on outdoor Gait3D, paralleling the finding that localized feature extraction is not consistently beneficial in wild data.The local branch is characterized as a 3D-convolutional analogue of FConv, whereas the global branch uses standard 3D convolution.
- SMPLGait: SMPLGait’s SMPL branch provides no apparent promotion when the silhouette branch uses a strong ResNet-like backbone.Possible reasons include sparse SMPL feature characterization, non-recognition-oriented modeling, and difficulty estimating accurate SMPL geometry from a single outdoor RGB camera.
- Analysis and Discussion: The paper argues that outdoor evaluation, comprehensive ablations, and stronger backbones are necessary because indoor datasets and plain convolutional networks may not reflect real-world complexity.It identifies indoor camera-array settings, limited dataset coverage, and weak backbones as important boundaries of earlier evaluations.
5. A Strong Baseline: GaitBase
GaitBase is a silhouette-based baseline built from simple, widely used components and designed for comparable indoor and outdoor evaluation. Its pipeline combines a ResNet-like backbone, temporal and horizontal pooling, and gait-specific augmentation.
- 5. A Strong Baseline: GaitBase: GaitBase is proposed as a structurally simple, experimentally powerful, and empirically robust silhouette-based baseline for indoor and outdoor evaluations.The model is intended to provide comparable or better performance without elaborate modules.
- Architecture: The pipeline follows set-based and part-based paradigms, using a ResNet-like backbone to transform silhouette frames into 3D feature maps.The architecture replaces the shallow convolutional backbone common in earlier gait-recognition systems with a residual design.
- Architecture: Temporal pooling maximizes across the sequence dimension to produce a set-level representation, while horizontal pooling extracts part-level information without a pyramid structure.The pipeline labels these operations TP and HP, respectively.
- Data Augmentation: GaitBase uses silhouette-specific spatial augmentation and variable-length training sequences to improve recognition accuracy and better match practical usage.The explored spatial operations include flipping, rotation, perspective, and affine transformations.
- Design Philosophy: The design philosophy favors succinct, widespread, and validated techniques rather than unique modules while targeting robust performance across indoor and outdoor datasets.The selected techniques include set-based sequence treatment, part-based modeling, a ResNet-like backbone, and general data augmentation.
6. Experiment
GaitBase is evaluated against published methods across indoor and outdoor datasets, with particularly strong margins on practical-scene datasets. Ablations further examine BNNeck, data augmentation, and backbone complexity.
- GaitBase achieves competitive or better performance than state-of-the-art methods on all compared datasets except CASIA-B.
- +9.5% on GREW and +18.3% on Gait3D are GaitBase’s margins over other methods in practical scenarios.
- CASIA-B produces less competitive accuracy, while re-segmented CASIA-B* retains relatively high performance.
- BNNeck generally improves GaitBase across almost all datasets.
- Data augmentation improves outdoor performance by +2.4% on GREW and +9.9% on Gait3D.
- ResNet50 adds over 3× computation cost while causing over-fitting on CASIA-B* and only +1.9% improvement on Gait3D.
7. Conclusion
The paper releases OpenGait, re-evaluates gait recognition methods, and proposes GaitBase as a simple, efficient baseline. GaitBase achieves the best performance on most datasets, especially those in the wild.
- OpenGait provides a fair, easy-to-use platform for implementing and comparing gait recognition methods efficiently.
- Fair ablation studies reveal insights that differ from conclusions in some original papers.
- GaitBase achieves the best performance on most gait datasets, especially outdoor datasets.
8. Supplementary Material
Supplementary experiments show that augmentation and random input lengths affect indoor and outdoor datasets differently. They also expose practical limits and motivate stronger baselines, verification research, and unsupervised learning.
- Effect of Spatial Data Augmentation: Horizontal Flip significantly improves Gait3D but fails to improve CASIA-B*, reflecting their different viewpoint coverage.
- Effect of Spatial Data Augmentation: Rotation slightly benefits both CASIA-B* and Gait3D.
- Effect of Spatial Data Augmentation: Affine Transformation brings no performance gain on CASIA-B* or Gait3D because it does not effectively simulate their noisy factors.
- Effect of Spatial Data Augmentation: Random Erasing makes almost no difference on Gait3D because it poorly simulates practical occlusion cases.
- Effect of Spatial Data Augmentation: The selected augmentation strategy improves accuracy by 1.9% on CASIA-B* and 7.7% on Gait3D.
- Effect of Random Training Input Length: Fixed input lengths work relatively best indoors, whereas random lengths yield superior performance on GREW and Gait3D.
- Future Work and Discussion: GREW reaches 60.1% rank-1 accuracy, but a significant gap remains before gait recognition is accurate enough for real-world applications.
- Future Work and Discussion: Annotated outdoor gait data are economically expensive and limited in the trade-off between diversity and scale.
Ethical Statements.
The authors emphasize that improper use or abuse of gait recognition threatens personal privacy. They argue that vision technology should serve human happiness.
- Improper use or abuse of gait recognition may threaten personal privacy.
- The authors argue that vision-technique development should serve human happiness.