Research map / Label-efficient and robust learning / Recognition and long-tail
SkeletonGait: Gait Recognition Using Skeleton Maps
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 the skeleton map, a Gaussian-rendered heatmap of normalized joint coordinates, with SkeletonGait (DeepGaitV2 fed skeleton maps) and SkeletonGait++ (silhouette+skeleton two-branch fusion) for gait recognition.
Problem
Gait recognition relies on silhouettes that are unreliable in unconstrained scenes, while existing skeleton-based gait methods using raw joint coordinates with GCNs perform poorly on real-world datasets.
Approach
Render normalized pose joints and limbs as a silhouette-like Gaussian skeleton map (center- and scale-normalized, subject-cropped) so skeletons can be processed by image-based gait backbones, then fuse this branch with a silhouette branch.
Key results
- SkeletonGait reaches 77.4% rank-1 on GREW and 38.1% rank-1 on Gait3D, outperforming prior skeleton-based methods.
- SkeletonGait++ achieves 85.8% rank-1 on GREW, 77.6% on Gait3D, and 81.3% overall rank-1 on SUSTech1K.
- On SUSTech1K night condition, SkeletonGait (43.9%) surpasses DeepGaitV2 (27.3%) by a large margin.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace silhouette input with skeleton maps | DeepGaitV2, yielding SkeletonGait · GREW | rank-1 accuracy: 77.7 -> 77.4 | No clear effect |
| Add a skeleton-map branch to the silhouette model | DeepGaitV2, yielding SkeletonGait++ · GREW | rank-1 accuracy: +8.1% | Helped |
| Add a skeleton-map branch to the silhouette model | DeepGaitV2, yielding SkeletonGait++ · Gait3D | rank-1 accuracy: +3.2% | Helped |
| Add a skeleton-map branch to the silhouette model | DeepGaitV2, yielding SkeletonGait++ · SUSTech1K | rank-1 accuracy: +5.2% | Helped |
| Set skeleton-map Gaussian σ to 1.0 instead of the default 8.0 (ablation) | SkeletonGait · Gait3D | rank-1 accuracy: 38.1 -> 34.3 | Hurt |
| Set skeleton-map Gaussian σ to 4.0 instead of the default 8.0 (ablation) | SkeletonGait · Gait3D | rank-1 accuracy: 38.1 -> 37.5 | Hurt |
| Set skeleton-map Gaussian σ to 16.0 instead of the default 8.0 (ablation) | SkeletonGait · Gait3D | rank-1 accuracy: 38.1 -> 36.0 | Hurt |
| Use low-level attention fusion instead of add fusion (ablation) | SkeletonGait++ · Gait3D | rank-1 accuracy: 76.5 -> 77.6 | Helped |
| Use low-level attention fusion instead of concatenate fusion (ablation) | SkeletonGait++ · Gait3D | mAP: 69.7 -> 70.3 | Helped |
| Use high-level attention fusion instead of add fusion (ablation) | SkeletonGait++ · Gait3D | rank-1 accuracy: 76.2 -> 78.2 | Helped |
| Use high-level attention fusion instead of concatenate fusion (ablation) | SkeletonGait++ · Gait3D | rank-1 accuracy: 76.5 -> 78.2 | Helped |
Surprising findings
- Skeleton-only SkeletonGait nearly matches silhouette-based DeepGaitV2 on GREW, which the authors attribute to limited viewpoint diversity in GREW's test pairs.
- Cross-view changes remain a major challenge for skeleton-based methods.
Limitations
- Cross-view recognition remains difficult for skeleton-based methods.
- Skeleton data can become unreliable under extensive occlusion or challenging conditions.
- OU-MVLP results for SkeletonGait++ are absent due to lack of frame-by-frame alignment between skeleton and silhouette.
Experimental setup
Datasets: OU-MVLP GREW Gait3D SUSTech1K CCPG
Benchmarks: OU-MVLP GREW Gait3D CCPG SUSTech1K
Metrics: rank-1 rank-5 mAP mINP
Baselines: GaitGraph2 GaitTR GPGait GaitSet GaitPart GaitGL GaitBase DeepGaitV2 SMPLGait GaitRef BiFusion AUG-OGBase
Builds on: DeepGaitV2 OpenGait GaitGraph GaitSet
Closest papers by meaning
- Motif Guided Graph Transformers with Combinatorial Skeleton Prototype Learning for Skeleton-Based Person Re-Identification (AAAI 2025 · similarity 0.86)
- Exploring More from Multiple Gait Modalities for Human Identification (AAAI 2025 · similarity 0.86)
- Vocabulary-Guided Gait Recognition (NeurIPS 2025 · similarity 0.86)
- BigGait: Learning Gait Representation You Want by Large Vision Models (CVPR 2024 · similarity 0.86)
- Revealing Key Details to See Differences: A Novel Prototypical Perspective for Skeleton-based Action Recognition (CVPR 2025 · similarity 0.85)
- QAGait: Revisit Gait Recognition from a Quality Perspective (AAAI 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.