Research map / Label-efficient and robust learning / Recognition and long-tail
RA-GAR: A Richly Annotated Benchmark for Gait Attribute 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.
Introduces RA-GAR, a 533-subject, 123k-sequence gait attribute benchmark with 15 attributes under covariates, and CLIP-GAR, a two-stage CLIP-aligned gait attribute recognizer.
Problem
Gait attribute recognition is limited by existing datasets that have few attributes and ignore covariate scenarios such as carrying bags and clothing changes; existing methods also lack attribute-specific local perception and semantic correlation modeling.
Approach
Align a gait-specific silhouette encoder (DeepGaitV2-like) with frozen CLIP text embeddings of templated attribute sentences via a contrastive KL loss, then fuse frozen visual part tokens and text tokens with a Transformer block for per-attribute classification.
Key results
- CLIP-GAR reaches 65.57% mA and 76.00% F1 on RA-GAR, versus 62.18% mA for GAR-Net (+3.39 mA) and 73.15 F1 (+2.85 F1 reported in text).
- On MA-Gait, CLIP-GAR achieves 74.69% mA and 77.64% F1, improving mA by 2.83 points and F1 by 0.84 points over GAR-Net.
- Zero-shot CLIP ResNet-50 achieves only 33.97% F1 on RA-GAR, motivating the gait-specific visual encoder.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace the CLIP ResNet-50 visual encoder with a gait-specific DeepGaitV2-3D encoder and add Align and Fusion stages (ablation) | CLIP-GAR · RA-GAR | F1: 71.48 -> 76.00 | Helped |
| Replace the CLIP ResNet-50 visual encoder with a gait-specific DeepGaitV2-3D encoder and add Align and Fusion stages (ablation) | CLIP-GAR · RA-GAR | mA: 71.48 -> 76.00 | Helped |
| Use the zero-shot CLIP ResNet-50 visual encoder (ablation) | CLIP-GAR · RA-GAR | F1: 33.97 | Hurt |
| Fine-tune the CLIP ResNet-50 visual encoder instead of using it zero-shot (ablation) | CLIP-GAR · RA-GAR | F1: 33.97 -> 61.85 | Helped |
| Add the Align stage to DeepGaitV2-3D classification (ablation) | CLIP-GAR · RA-GAR | F1: 71.48 -> 73.04 | Helped |
| Add the Fusion stage after Align (ablation) | CLIP-GAR · RA-GAR | F1: 73.04 -> 76.00 | Helped |
| Use 16 rather than 8 horizontal parts (ablation) | CLIP-GAR · RA-GAR | F1: 73.04 vs. 72.59; 76.00 vs. 75.61 | Hurt |
| Use 4 rather than 8 horizontal parts (ablation) | CLIP-GAR · RA-GAR | F1: 73.04 vs. 73.67; 76.00 vs. 75.09 | Mixed |
| Use 1 part (no horizontal part division) rather than 8 parts (ablation) | CLIP-GAR · RA-GAR | F1: 73.04 vs. 71.82; 76.00 vs. 73.25 | Hurt |
| Use 4 rather than 8 horizontal parts (ablation) | CLIP-GAR · RA-GAR | mA: 62.57 vs. 62.67; 65.57 vs. 66.37 | Mixed |
| Use 1 part (no horizontal part division) rather than 8 parts (ablation) | CLIP-GAR · RA-GAR | mA: 62.57 vs. 61.79; 65.57 vs. 64.17 | Hurt |
| Use CLIP-GAR rather than GAR-Net | gait attribute recognition · RA-GAR | F1: 73.15 -> 76.00 | Helped |
Surprising findings
- Zero-shot CLIP (ResNet-50) reaches only 33.97% F1 on gait attribute recognition, showing a large domain gap from RGB images to silhouettes.
- Adding ground-truth attribute labels to gait features substantially raises identity recognition Rank-1 for most methods (e.g., GaitBase 86.14 to 99.54).
Limitations
- Self-occlusion makes some fine-grained attributes hard to predict.
- Class imbalance across attributes remains a challenge.
- Attributes are partly simulated by subjects following instructions rather than naturally occurring.
Experimental setup
Datasets: RA-GAR MA-Gait
Benchmarks: RA-GAR MA-Gait
Metrics: mA F1 precision recall accuracy
Baselines: GaitGraph GaitGraph2 GaitTR GPGait GaitSet GaitPart GaitGL GaitBase DeepGaitV2 GAR-Net
Builds on: CLIP DeepGaitV2 GAR-Net
Closest papers by meaning
- Cross-Covariate Gait Recognition: A Benchmark (AAAI 2024 · similarity 0.90)
- Exploring More from Multiple Gait Modalities for Human Identification (AAAI 2025 · similarity 0.86)
- Vocabulary-Guided Gait Recognition (NeurIPS 2025 · similarity 0.86)
- BigGait: Learning Gait Representation You Want by Large Vision Models (CVPR 2024 · similarity 0.85)
- BiggerGait: Unlocking Gait Recognition with Layer-wise Representations from Large Vision Models (NeurIPS 2025 · similarity 0.85)
- Bridging Gait Recognition and Large Language Models Sequence Modeling (CVPR 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.