Research map / Vision-language and multimodal / Human and embodied modeling
Learning Visual Prompt 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 VPNet, a gait recognition network combining a temporal-shift 3D-conv ResNet backbone with a trainable part-based visual prompt pool and a dynamic transformer with a context window mask, evaluated on Gait3D, GREW, OUMVLP and CASIA-B.
Problem
Gait recognition in real-world scenarios is hard due to cross-view, cross-clothing variation, occlusions, and turning or lingering, and existing CNN methods learn fixed patterns and suffer from length generalization when test sequences are longer than training subsequences.
Approach
Introduce a trainable part-based prompt pool of key-value pairs that dynamically selects task-relevant prompt templates to concatenate with gait sequence features, combined with a dynamic attention that uses a context window mask to handle length generalization, plus a forward cyclic temporal shift module for global motion patterns.
Key results
- VPNet-L reaches 75.4% rank-1 on Gait3D, a 9.1-point gain over DyGait (66.3%).
- VPNet-L reaches 80.0% rank-1 on GREW, outperforming GaitGCI (68.5%) by 11.5 points.
- VPNet-T achieves 94.9% mean rank-1 on CASIA-B, and VPNet-M reaches 92.4% mean rank-1 on OUMVLP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add FTS (ablation) | VPNet · Gait3D | rank-1 accuracy: 73.2 -> 75.4 | Helped |
| add the prompt pool (ablation) | VPNet · Gait3D | rank-1 accuracy: 72.8 -> 75.4 | Helped |
| add the context window mask (ablation) | VPNet · Gait3D | rank-1 accuracy: 73.0 -> 75.4 | Helped |
| use dynamic attention instead of max+mean pooling (ablation) | VPNet · Gait3D | rank-1 accuracy: 72.1 -> 75.4 | Helped |
| add FTS to 2D convolution (ablation) | VPNet temporal shift bottleneck · Gait3D | rank-1 accuracy: 69.7 -> 74.1 | Helped |
| add FTS to 3D convolution (ablation) | VPNet temporal shift bottleneck · Gait3D | rank-1 accuracy: 73.2 -> 75.4 | Helped |
| replace 3D convolution with 2D convolution, without FTS (ablation) | VPNet temporal shift bottleneck · Gait3D | rank-1 accuracy: 73.2 -> 69.7 | Hurt |
| use local length 8 instead of the full method's length (ablation) | VPNet dynamic attention · Gait3D | rank-1 accuracy: 75.4 -> 74.9 | Hurt |
| use local length 16 instead of the full method's length (ablation) | VPNet dynamic attention · Gait3D | rank-1 accuracy: 73.2 -> 75.4 | Helped |
| use local length 32 instead of the full method's length (ablation) | VPNet dynamic attention · Gait3D | rank-1 accuracy: 75.4 -> 73.2 | Hurt |
Surprising findings
- Replacing the 3D convolution with 2D convolution in the temporal-shift bottleneck drops Gait3D rank-1 to 69.7%, while adding FTS to 2D conv raises it to 74.1%.
Limitations
- The network does not use large-scale pre-training, unlike prompt learning in NLP.
- Larger local receptive fields (l=32) introduce noise and reduce Gait3D accuracy.
- Indoor lab benchmarks are near saturation, so the authors focus on outdoor scenarios.
Experimental setup
Datasets: Gait3D GREW OUMVLP CASIA-B
Benchmarks: Gait3D GREW OUMVLP CASIA-B
Metrics: rank-1 accuracy rank-5 accuracy
Baselines: GaitSet GaitPart GLN GaitGL GaitBase GaitGCI DANet HSTL DyGait SMPLGait
Builds on: ResNet temporal shift module VPT BNNeck VQ-VAE
Closest papers by meaning
- Dynamic Aggregated Network for Gait Recognition (CVPR 2023 · similarity 0.90)
- BiggerGait: Unlocking Gait Recognition with Layer-wise Representations from Large Vision Models (NeurIPS 2025 · similarity 0.89)
- Vocabulary-Guided Gait Recognition (NeurIPS 2025 · similarity 0.88)
- BigGait: Learning Gait Representation You Want by Large Vision Models (CVPR 2024 · similarity 0.88)
- Bridging Gait Recognition and Large Language Models Sequence Modeling (CVPR 2025 · similarity 0.86)
- On Denoising Walking Videos for 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.