Research map / Architectures and efficiency / RNNs and normalization
BiggerGait: Unlocking Gait Recognition with Layer-wise Representations from 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 BiggerGait, a layer-wise gait recognition baseline that extracts gait representations from every (or grouped) intermediate layer of frozen-or-tuned large vision models (DINOv2, CLIP, SAM) and fuses them, with only a mask auto-encoder as gait prior.
Problem
Existing LVM-based gait recognition relies heavily on hand-crafted gait priors and uses only the final LVM layer, underexploiting the rich representations across intermediate layers.
Approach
Treat each LVM layer as a separate gait representation source: use a lightweight auto-encoder silhouette mask as the only gait prior, attach a per-layer gait encoder, and fuse layer outputs at test time via ensemble distance; a grouping variant (BiggerGait*) shares encoders across depth groups to cut cost.
Key results
- Ensemble testing across layers raises accuracy by +7.7% on DINOv2-S, +10.3% on CLIP-S, and +7.0% on SAM-S on CCPG (per the layer analysis).
- Trained on CCPG, DINOv2-S-based BiggerGait reaches 67.0 average accuracy across the test sets versus 53.0 for BigGait (DINOv2-S), with 93.0 CL on SUSTech1K.
- BiggerGait* with DINOv2-S (2 gait encoders, depth groups J=6) cuts FLOPs to 29.8G from 45.7G while keeping average accuracy near 66.5.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Fuse layer-wise representations using ensemble testing instead of relying on a single layer | BiggerGait · CCPG and cross-domain gait evaluation | rank-1 accuracy: +7.7% on DINOv2-S, +10.3% on CLIP-S, and +7.0% on SAM-S | Helped |
| Use intermediate layers rather than the final layer | LVM gait representations · CCPG | average rank-1 accuracy: 89.5% ... Layer 6; ... final layer ... 80.9% | Helped |
| Use ensemble testing across layers | BiggerGait · CCGR_MINI | rank-1 accuracy | Hurt |
| Use P=2 shared gait encoders instead of P=12 (ablation) | BiggerGait* with DINOv2-S · CCPG | rank-1 accuracy (CL): 89.8 -> 89.5 | No clear effect |
| Use P=1 shared gait encoder instead of P=12 (ablation) | BiggerGait* with DINOv2-S · CCPG | rank-1 accuracy (CL): 89.8 -> 86.3 | Hurt |
| Use J=6 depth groups instead of J=12 (ablation) | BiggerGait* with DINOv2-S · CCPG | rank-1 accuracy (CL): 89.5 -> 89.0 | No clear effect |
| Use J=3 depth groups instead of J=12 (ablation) | BiggerGait* with DINOv2-S · CCPG | rank-1 accuracy (CL): 89.5 -> 88.8 | Hurt |
| Apply depth and gait grouping to larger LVMs (ablation) | BiggerGait with SAM-L, CLIP-L, and DINOv2-L · CCPG | rank-1 accuracy (CL): 90.0 -> 89.0; 87.5 -> 85.6; 92.8 -> 92.7 | Hurt |
| Use layer-wise representations instead of the single-level BigGait design | DINOv2-S gait recognition · CCPG train; multi-dataset evaluation | average rank-1 accuracy: 53.0 -> 67.0 | Helped |
| Use layer-wise representations instead of the single-level BigGait design | DINOv2-S gait recognition · CCGR_MINI train; multi-dataset evaluation | average rank-1 accuracy: 73.7 -> 78.1 | Helped |
Surprising findings
- Intermediate layers (e.g., layer 6 of DINOv2-S) outperform the final layer, with the final layer dropping to 80.9% from 89.5% average accuracy.
- Even a single shared gait encoder across all layers still yields state-of-the-art results for DINOv2-S and SAM-S based BiggerGait.
Limitations
- Feature extraction is mainly image-level; temporal modeling is underexplored.
- Layer-wise complementarity is weak on CCGR_MINI, where ensemble testing sometimes drops performance and BiggerGait underperforms BigGait within-domain.
- Grouping offers marginal FLOP savings and slight performance loss for larger LVMs (L variants).
Experimental setup
Datasets: CCPG CCGR_MINI
Benchmarks: CCPG CCGR_MINI SUSTech1K CASIA-B*
Metrics: Rank-1 accuracy Rank-5 accuracy parameters FLOPs
Baselines: GaitSet GaitPart GaitGL GaitBase DeepGaitV2 BiFusion SkeletonGait++ XGait MultiGait++ GaitEdge BigGait
Builds on: BigGait GaitBase DINOv2 CLIP SAM
Closest papers by meaning
- BigGait: Learning Gait Representation You Want by Large Vision Models (CVPR 2024 · similarity 0.94)
- Bridging Gait Recognition and Large Language Models Sequence Modeling (CVPR 2025 · similarity 0.89)
- Learning Visual Prompt for Gait Recognition (CVPR 2024 · similarity 0.89)
- Vocabulary-Guided Gait Recognition (NeurIPS 2025 · similarity 0.87)
- Cross-Covariate Gait Recognition: A Benchmark (AAAI 2024 · similarity 0.86)
- LidarGait++: Learning Local Features and Size Awareness from LiDAR Point Clouds for 3D Gait Recognition (CVPR 2025 · 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.