Research map / 3D vision / 3D generation
HybridGait: A Benchmark for Spatial-Temporal Cloth-Changing Gait Recognition with Hybrid Explorations
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 CCGait, an in-the-wild cloth-changing gait benchmark captured over 92 days, and HybridGait, a three-branch SMPL-aided model combining silhouettes, a CA-STT temporal branch over 3D mesh dynamics, and a SilD projection branch.
Problem
Gait recognition under persistent clothing and appearance changes over time and space in in-the-wild settings, plus the coupling of clothing and viewpoint variations, which existing lab-based benchmarks do not capture.
Approach
Combines silhouette appearance features with 3D SMPL mesh temporal dynamics (via a canonical-aligned spatial-temporal transformer) and 3D-to-2D projected silhouettes refined by silhouette-guided deformable alignment, fusing the three branches to handle clothing and viewpoint changes.
Key results
- HybridGait reaches 51.12% Rank-1 and 51.30% mAP on CCGait, versus 48.60% Rank-1 for the next best (SMPLGait).
- On Gait3D, HybridGait achieves 53.30% Rank-1 and 43.29% mAP, improving over GaitGCI's 50.30% Rank-1 and SMPLGait's 46.30%.
- Ablation on Gait3D shows the full model (Appr + CA-STT + SilD) raises Rank-1 from 42.90% (appearance only) to 53.30%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use HybridGait instead of SMPLGait | CCGait gait recognition · CCGait | Rank-1: 48.60 -> 51.12 | Helped |
| use HybridGait instead of GaitGCI | Gait3D gait recognition · Gait3D | Rank-1: 50.30 -> 53.30 | Helped |
| add a temporal branch with STT to the appearance branch (ablation) | HybridGait · Gait3D | Rank-1: 42.90 -> 48.90 | Helped |
| add a temporal branch with STT to the appearance branch (ablation) | HybridGait · CCGait | Rank-1: 44.31 -> 47.36 | Helped |
| replace STT with CA-STT (ablation) | HybridGait temporal branch · Gait3D | Rank-1: 48.90 -> 50.26 | Helped |
| replace STT with CA-STT (ablation) | HybridGait temporal branch · CCGait | Rank-1: 47.36 -> 50.28 | Helped |
| add SilD to the appearance branch (ablation) | HybridGait · Gait3D | Rank-1: 42.90 -> 46.31 | Helped |
| add SilD to the appearance branch (ablation) | HybridGait · CCGait | Rank-1: 44.31 -> 46.66 | Helped |
| add SilD to Appr + STT (ablation) | HybridGait · Gait3D | Rank-1: 48.90 -> 52.10 | Helped |
| add SilD to Appr + STT (ablation) | HybridGait · CCGait | Rank-1: 47.36 -> 50.35 | Helped |
| add SilD to Appr + CA-STT (ablation) | HybridGait · Gait3D | Rank-1: 50.26 -> 53.30 | Helped |
| add SilD to Appr + CA-STT (ablation) | HybridGait · CCGait | Rank-1: 50.28 -> 51.12 | Helped |
Surprising findings
- The model-based GaitGraph performs very poorly on CCGait (about 7% Rank-1), far below silhouette-based methods.
Limitations
- RGB images are withheld for privacy, so the dataset cannot be fully released in raw form.
- MTSGait and GaitGCI could not be evaluated on CCGait because their code is not publicly available.
Experimental setup
Datasets: CCGait Gait3D
Benchmarks: CCGait Gait3D
Metrics: Rank-1 Rank-5 mAP mINP
Baselines: GaitGraph GEINet GaitSet GaitPart GaitGL CSTL MTSGait GaitGCI SMPLGait
Builds on: SMPLGait GaitSet deformable convolution Horizontal Pyramid Pooling
Closest papers by meaning
- SkeletonGait: Gait Recognition Using Skeleton Maps (AAAI 2024 · similarity 0.85)
- Exploring More from Multiple Gait Modalities for Human Identification (AAAI 2025 · similarity 0.85)
- BigGait: Learning Gait Representation You Want by Large Vision Models (CVPR 2024 · similarity 0.84)
- Learning Visual Prompt for Gait Recognition (CVPR 2024 · similarity 0.84)
- QAGait: Revisit Gait Recognition from a Quality Perspective (AAAI 2024 · similarity 0.84)
- Cross-Covariate Gait Recognition: A Benchmark (AAAI 2024 · similarity 0.84)
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.