Research map / Label-efficient and robust learning / Recognition and long-tail
BigGait: Learning Gait Representation You Want by Large Vision Models
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 BigGait, which replaces task-specific supervised upstream models in gait recognition with a frozen self-supervised DINOv2 and a Gait Representation Extractor (mask, appearance, denoising branches) feeding a modified GaitBase.
Problem
Gait recognition relies on task-specific supervised upstream models (segmentation, parsing, pose, SMPL) that are costly to annotate and introduce error accumulation, restricting gait representations to predetermined forms.
Approach
Derive implicit gait representations directly from all-purpose features of a frozen self-supervised large vision model (DINOv2) using unsupervised mask learning, a parsing-like appearance branch, and a denoising branch with smoothness and diversity soft constraints, instead of supervised silhouettes or parsing.
Key results
- BigGait reaches 87.2% mean rank-1 on CCPG, versus 83.7% for SkeletonGait++ and 83.3% for DeepGaitV2 (best reported).
- On CCPG, BigGait beats video-based ReID methods by +29.2% on CL, +28.6% on UP, +17.4% on DN, and +1.7% on BG.
- Cross-domain results are mixed: trained on CCPG, BigGait reaches 56.4% overall on SUSTech1K versus 40.6% for PSTA, but trained on CASIA-B* it is weak on CL (7.5%) and DN (14.2%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use BigGait instead of SkeletonGait++ | CCPG · CCPG | rank-1 accuracy (mean): 83.7 -> 87.2 | Helped |
| Add Pad-and-Resize (ablation) | BigGait · CCPG | rank-1 accuracy (CL): 70.5 -> 76.0 | Helped |
| Add the GRE module (ablation) | BigGait · CCPG | rank-1 accuracy (CL): 55.6 -> 76.0 | Helped |
| Add the mask branch (ablation) | BigGait · CCPG | rank-1 accuracy (CL): 70.5 -> 76.0 | Helped |
| Add the denoising branch (ablation) | BigGait · CCPG | rank-1 accuracy (CL): 66.1 -> 76.0 | Helped |
| Add the appearance branch (ablation) | BigGait · CCPG | rank-1 accuracy (CL): 65.1 -> 76.0 | Helped |
| Use the GRE branches instead of feeding the foreground mask directly (ablation) | BigGait · CCPG | rank-1 accuracy (CL): 36.7 -> 76.0 | Helped |
| Add smoothness and diversity losses to the denoising branch (ablation) | BigGait · CCPG | rank-1 accuracy (CL): 69.1 -> 76.0 | Helped |
| Add the diversity loss (ablation) | BigGait denoising branch with smoothness loss · CCPG | rank-1 accuracy (CL): 68.8 -> 76.0 | Helped |
| Add the smoothness loss (ablation) | BigGait denoising branch with diversity loss · CCPG | rank-1 accuracy (CL): 74.5 -> 76.0 | Helped |
| Increase denoising branch dimension from 8 to 16 (ablation) | BigGait · CCPG | rank-1 accuracy (CL): 72.8 -> 76.0 | Helped |
| Reduce denoising branch dimension from 32 to 16 (ablation) | BigGait · CCPG | rank-1 accuracy (CL): 68.5 -> 76.0 | Helped |
Surprising findings
- Cross-domain performance depends heavily on training-data clothing-change distribution; BigGait struggles when trained on CASIA-B* or SUSTech1K, which contain few cross-dressing pairs.
- Fine-tuning DINOv2 hurts performance (74.8 CL vs 76.0 frozen), and DINOv2 from scratch performs far worse (36.6 CL).
Limitations
- Influence of different upstream LVMs only preliminarily explored.
- GRE module lacks spatial-temporal design.
- Cross-domain performance degrades when training data has few cross-clothing pairs; focuses on high-frequency texture rather than color noise.
Experimental setup
Datasets: CCPG CASIA-B* SUSTech1K
Benchmarks: CCPG CASIA-B* SUSTech1K
Metrics: rank-1 accuracy
Baselines: GaitSet GaitBase GaitPart DeepGaitV2 SkeletonGait++ GaitEdge AP3D PSTA PiT Gait-TR
Builds on: DINOv2 GaitBase GaitSet SAM
Closest papers by meaning
- BiggerGait: Unlocking Gait Recognition with Layer-wise Representations from Large Vision Models (NeurIPS 2025 · similarity 0.94)
- Learning Visual Prompt for Gait Recognition (CVPR 2024 · similarity 0.88)
- Bridging Gait Recognition and Large Language Models Sequence Modeling (CVPR 2025 · similarity 0.87)
- Vocabulary-Guided Gait Recognition (NeurIPS 2025 · similarity 0.86)
- SkeletonGait: Gait Recognition Using Skeleton Maps (AAAI 2024 · similarity 0.86)
- RA-GAR: A Richly Annotated Benchmark for Gait Attribute Recognition (AAAI 2025 · similarity 0.85)
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.