Research map / Video / Video segmentation and tracking
Exploring More from Multiple Gait Modalities for Human Identification
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.
Compares silhouette, human parsing, and optical flow gait modalities under fair settings and proposes MultiGait++ with a C²Fusion module that fuses shared and distinct features across three branches.
Problem
Gait recognition studies of silhouette, human parsing, and optical flow are inconsistent in architecture and experimental settings, so their relative value and fusion strategies are unclear.
Approach
A unified DeepGaitV2-based framework compares unimodal and multimodal gait baselines across fusion locations and mechanisms, then C²Fusion uses cross-attention-derived common and difference attention maps to extract shared features and emphasize branch-specific features.
Key results
- MultiGait++ reaches 87.4% overall rank-1 on SUSTech1K, versus 80.9% for the silhouette-only DeepGaitV2 baseline.
- MultiGait++ achieves 87.6% mean rank-1 on CCPG, +4.3% over DeepGaitV2 per the paper's appendix comparison.
- On real-world datasets, MultiGait++ reaches 93.4% R-1 on GREW (silhouette+flow) and 85.4% R-1 on Gait3D (silhouette+parsing).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use C^2Fusion instead of naive three-modality fusion | MultiGait++ · SUSTech1K | R-1: 86.0 -> 87.4 | Helped |
| Use C^2Fusion instead of naive three-modality fusion | MultiGait++ · CCPG mean | R-1: 85.5 -> 87.6 | Helped |
| Use C^2Fusion instead of silhouette-flow baseline fusion | MultiGait++ s+f · GREW | R-1: 91.4 -> 93.4 | Helped |
| Use C^2Fusion instead of silhouette-parsing baseline fusion | MultiGait++ s+p · Gait3D | R-1: 83.0 -> 85.4 | Helped |
| Add parsing to silhouette-only input | MultiGait · SUSTech1K | R-1: 80.9 -> 85.8 | Helped |
| Fuse silhouette and flow at high level rather than input level | MultiGait s+f · SUSTech1K | R-1: 81.3 -> 83.9 | Helped |
| Add common-concern map m_co while retaining m_di (ablation) | MultiGait++ · SUSTech1K | R-1: 86.5 -> 87.4 | Helped |
| Add different-concern map m_di while retaining m_co (ablation) | MultiGait++ · SUSTech1K | R-1: 86.4 -> 87.4 | Helped |
| Add both m_co and m_di rather than neither (ablation) | MultiGait++ · SUSTech1K | R-1: 86.0 -> 87.4 | Helped |
| Add skeleton map concatenated with silhouette and parsing (ablation) | MultiGait++ · SUSTech1K | R-1: 87.4 -> 87.8 | Helped |
Surprising findings
- Human parsing alone underperforms silhouettes (77.3% vs 80.9% R-1 on SUSTech1K), attributed to noise from fine-grained segmentation.
- Optical flow alone achieves competitive performance with silhouettes (80.5% vs 80.9% R-1 on SUSTech1K).
Limitations
- Parsing and optical flow must be extracted by external models, adding pipeline cost.
- Parsing and flow are not available in GREW and Gait3D, so the multimodal design is adapted to two-branch inputs there.
- Naive skeleton-map addition gave limited gain, suggesting more sophisticated integration is needed.
Experimental setup
Datasets: SUSTech1K CCPG GREW Gait3D Pascal-Person-Part
Benchmarks: SUSTech1K CCPG GREW Gait3D
Metrics: rank-1 accuracy rank-5 accuracy mAP mINP
Baselines: GaitSet GaitPart GaitGL GaitBase DeepGaitV2 SkeletonGait SkeletonGait++ BiFusion XGait GaitFusion BigGait GaitEdge
Builds on: DeepGaitV2 OpenGait SkeletonGait++ VideoFlow QANet
Closest papers by meaning
- QAGait: Revisit Gait Recognition from a Quality Perspective (AAAI 2024 · similarity 0.88)
- On Denoising Walking Videos for Gait Recognition (CVPR 2025 · similarity 0.87)
- Cross-Covariate Gait Recognition: A Benchmark (AAAI 2024 · similarity 0.87)
- SkeletonGait: Gait Recognition Using Skeleton Maps (AAAI 2024 · similarity 0.86)
- RA-GAR: A Richly Annotated Benchmark for Gait Attribute Recognition (AAAI 2025 · similarity 0.86)
- Learning Visual Prompt for Gait Recognition (CVPR 2024 · similarity 0.86)
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.