Research map / Label-efficient and robust learning / Recognition and long-tail
Cross-Covariate Gait Recognition: A Benchmark
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces CCGR, a 970-subject, ~1.6M-sequence gait benchmark with 53 covariates and 33 views plus RGB/parsing/silhouette/pose data, and ParsingGait, which runs silhouette-based gait models on human parsing maps.
Problem
Existing gait benchmarks, both constrained indoor and in-the-wild outdoor, lack covariate diversity at population and individual levels, so current methods look near-saturated and fail to reflect real-world cross-covariate challenges.
Approach
Collect a large, controlled, multi-view dataset in which every subject walks under many covariates (single and mixed) across 33 cameras, and use human parsing maps as a gait modality that can be fed directly into existing silhouette-based gait models without architectural changes.
Key results
- DeepGaitV2 achieves 42.5% hard rank-1 on CCGR versus about 82% on GREW, showing the benchmark is much harder.
- ParsingGait with DeepGaitV2 backbone reaches 52.7% hard rank-1, up from 42.5% for silhouette-based DeepGaitV2, and ParsingGait with GaitBase reaches 43.2% vs 31.3%.
- Using only population-level diversity drops accuracy by about 18.6% on average, and adding individual-level diversity drops it a further about 25.1%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use parsing instead of silhouette input with the DeepGaitV2 backbone | DeepGaitV2 · CCGR hard | Rank-1 accuracy: 42.5 -> 52.7 | Helped |
| Use parsing instead of silhouette input with the DeepGaitV2 backbone | DeepGaitV2 · CCGR easy | Rank-1 accuracy: 55.2 -> 67.2 | Helped |
| Increase covariate diversity from Group A to Groups B1/B2/B3 | CCGR evaluation setup · CCGR | accuracy: -18.6% | Hurt |
| Add individual-level diversity, moving from Groups B1/B2/B3 to C | CCGR evaluation setup · CCGR | accuracy: -25.1% | Hurt |
| Increase the number of covariates | Gait recognition models evaluated on CCGR · CCGR | accuracy | Hurt |
| Add more simultaneous covariates to mixed-covariate cases | Gait recognition models evaluated on CCGR · CCGR mixed-covariate evaluation | Rank-1 accuracy | Hurt |
Surprising findings
- Individual-level covariate diversity is as challenging as population-level diversity, not just a secondary factor.
- GaitGraph2 scores near zero (0.26% hard rank-1), and GEINet scores 3.10%, far below silhouette-based methods.
Limitations
- Paper is written as a dataset release; the text says CCGR 'will be made publicly available', so release status is uncertain.
- Evaluation relies on the authors' own re-implementations of baselines with fixed protocols.
- Parsing extraction depends on QANet, and the paper reports that alternative segmentation methods perform worse.
Experimental setup
Datasets: CCGR CASIA-B OU-MVLP GREW Gait3D
Benchmarks: CCGR
Metrics: rank-1 accuracy rank-5 accuracy
Baselines: GEINet GaitSet GaitPart GaitGL CSTL GaitBase DeepGaitV2 GaitGraph GaitGraph2
Builds on: GaitSet GaitPart GaitGL GaitBase DeepGaitV2 QANet
Closest papers by meaning
- RA-GAR: A Richly Annotated Benchmark for Gait Attribute Recognition (AAAI 2025 · similarity 0.90)
- Exploring More from Multiple Gait Modalities for Human Identification (AAAI 2025 · similarity 0.87)
- Vocabulary-Guided Gait Recognition (NeurIPS 2025 · similarity 0.87)
- BiggerGait: Unlocking Gait Recognition with Layer-wise Representations from Large Vision Models (NeurIPS 2025 · similarity 0.86)
- On Denoising Walking Videos for Gait Recognition (CVPR 2025 · similarity 0.86)
- Learning Visual Prompt for Gait Recognition (CVPR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.