Research map / Video / Video segmentation and tracking
On Denoising Walking Videos for Gait Recognition
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.
Proposes DenoisingGait, which combines one-step Stable Diffusion feature denoising with a geometry-driven feature matching module that turns diffusion features into static and dynamic 2-channel Gait Feature Fields for RGB-based gait recognition.
Problem
Vision-based gait recognition must remove identity-irrelevant cues (clothing texture, color) from RGB walking videos while retaining identity-discriminative gait patterns; silhouette and pose inputs are sparse and RGB methods remain noisy.
Approach
Use one-step denoising features from a pretrained latent diffusion model (Stable Diffusion 1.5) as knowledge-driven denoising, then condense multi-channel diffusion features at each foreground pixel into a two-dimensional direction vector via within-frame (static) and cross-frame (dynamic) feature matching, yielding Gait Feature Fields; texture suppression via random zero-padding of high-magnitude
Key results
- On CCPG within-domain gait protocol, DenoisingGait reaches 89.5% mean rank-1 versus 87.2% for BigGait (CVPR'24).
- Cross-domain (trained on CCPG, tested on CASIA-B* and SUSTech1K) overall rank-1 of 59.1 versus 56.4 for BigGait.
- Ablation: diffusion features give 89.5% vs 83.1% without diffusion (RGB input 72.2% on CCPG mean with feature matching disabled).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use diffusion features instead of RGB images (ablation) | DenoisingGait on CCPG · CCPG, mean R1 | Rank-1 accuracy: 85.8 -> 89.5 | Helped |
| use diffusion features instead of gray images (ablation) | DenoisingGait on CCPG · CCPG, mean R1 | Rank-1 accuracy: 83.8 -> 89.5 | Helped |
| use diffusion features instead of DINOv2 features (ablation) | DenoisingGait on CCPG · CCPG, mean R1 | Rank-1 accuracy: 84.9 -> 89.5 | Helped |
| use diffusion features instead of silhouette and flow (ablation) | DenoisingGait on CCPG · CCPG, mean R1 | Rank-1 accuracy: 82.8 -> 89.5 | Helped |
| add feature matching (ablation) | DenoisingGait on CCPG · CCPG, mean R1 | Rank-1 accuracy: 83.1 -> 89.5 | Helped |
| add within-frame matching (ablation) | DenoisingGait on CCPG · CCPG, mean R1 | Rank-1 accuracy: 77.7 -> 89.5 | Helped |
| add cross-frame matching (ablation) | DenoisingGait on CCPG · CCPG, mean R1 | Rank-1 accuracy: 88.5 -> 89.5 | Helped |
| add background removal (ablation) | DenoisingGait on CCPG · CCPG, mean R1 | Rank-1 accuracy: 79.0 -> 89.5 | Helped |
| add texture suppression (ablation) | DenoisingGait on CCPG · CCPG, mean R1 | Rank-1 accuracy: 88.2 -> 89.5 | Helped |
| use DenoisingGait instead of BigGait | CCPG gait evaluation · CCPG, gait evaluation UP | Rank-1 accuracy: 89.0 -> 88.0 | Hurt |
| use DenoisingGait instead of BigGait | SUSTech1K within-domain evaluation · SUSTech1K, R-1 | Rank-1 accuracy: 96.2 -> 95.4 | Hurt |
| use DenoisingGait instead of BigGait | SUSTech1K within-domain evaluation · SUSTech1K, NT | Rank-1 accuracy: 85.3 -> 69.5 | Hurt |
Surprising findings
- Diffusion features at timestep t=700 improve GaitBase by 5.3% over the no-denoising baseline, while DINOv2 features give no improvement over the RGB-based pipeline in the ablation.
- Texture suppression via random zero-padding of high-magnitude pixels improves CL accuracy by 3.0%.
Limitations
- Performance on night conditions in SUSTech1K is limited by low-quality silhouettes.
- Large vision models such as diffusion and DINOv2 demand substantial training data, compute and storage.
- Limited clothing-change diversity in most gait datasets constrains RGB-based generalization.
Experimental setup
Datasets: CCPG CASIA-B* SUSTech1K
Benchmarks: CCPG CASIA-B* SUSTech1K
Metrics: rank-1 accuracy rank-5 accuracy mean accuracy
Baselines: GaitSet GaitPart GaitBase DeepGaitV2 BigGait SkeletonGait++ MultiGait++ XGait GaitEdge BiFusion
Builds on: GaitBase Latent Diffusion Models BigGait SIFT optical flow
Closest papers by meaning
- Exploring More from Multiple Gait Modalities for Human Identification (AAAI 2025 · similarity 0.87)
- Learning Visual Prompt for Gait Recognition (CVPR 2024 · similarity 0.86)
- QAGait: Revisit Gait Recognition from a Quality Perspective (AAAI 2024 · similarity 0.86)
- Cross-Covariate Gait Recognition: A Benchmark (AAAI 2024 · similarity 0.86)
- BiggerGait: Unlocking Gait Recognition with Layer-wise Representations from Large Vision Models (NeurIPS 2025 · similarity 0.85)
- DeTrack: In-model Latent Denoising Learning for Visual Object Tracking (NeurIPS 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.