Research map / Label-efficient and robust learning / Recognition and long-tail
QAGait: Revisit Gait Recognition from a Quality Perspective
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 QAGait, a silhouette quality assessment pipeline (Maximal Connect Area, Template Match with Hu moments, lean-aware alignment) plus quality-adaptive margin CE and triplet losses for gait recognition, built on GaitBase.
Problem
Gait recognition on in-the-wild datasets suffers from unidentifiable low-quality silhouettes (segmentation errors, occlusion, non-human shapes) and identifiable but challenging silhouettes (background noise, non-standard posture) that current pipelines treat uniformly.
Approach
Cheap OpenCV-based quality filtering and lean-aware alignment before the backbone, plus a partial-feature-norm quality indicator that adaptively sets margins in CE and triplet losses, without network modification.
Key results
- On Gait3D (after quality assessment and quality-aware learning), QAGait reaches 67.0% Rank-1, a 7.3-point gain over GaitBase on the filtered Gait3D# (59.7%).
- On GREW#, QAGait achieves 59.1% Rank-1, improving 1.9 points over GaitBase on the filtered set (57.2%).
- On CASIA-B CL condition, QAGait with GaitBase backbone reaches 78.2% vs 75.1% original GaitBase.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the full QAGait framework | GaitBase · Gait3D# | Rank-1: 62.4 -> 67.0 | Helped |
| Add the full QAGait framework | GaitBase · GREW# | Rank-1: 58.9 -> 59.1 | Helped |
| Add the full QAGait framework | GaitBase · CASIA-B# CL | accuracy: 77.1 -> 78.2 | Helped |
| Add Maximal Connect quality assessment (ablation) | GaitBase on original Gait3D without data augmentation · Gait3D | Rank-1: 55.1 -> 55.8 | Helped |
| Add Template Match after Maximal Connect (ablation) | GaitBase on original Gait3D without data augmentation · Gait3D | Rank-1: 55.8 -> 54.9 | Hurt |
| Add data augmentation to quality assessment (ablation) | GaitBase on Gait3D · Gait3D | Rank-1: 54.9 -> 62.2 | Helped |
| Add alignment (ablation) | GaitBase with QA and data augmentation on Gait3D · Gait3D | Rank-1: 62.2 -> 63.5 | Helped |
| Add QACE (ablation) | GaitBase with QA, data augmentation, and alignment on Gait3D · Gait3D | Rank-1: 63.5 -> 66.0 | Helped |
| Add QATriplet (ablation) | GaitBase with QA, data augmentation, alignment, and QACE on Gait3D · Gait3D | Rank-1: 66.0 -> 67.0 | Helped |
| Restrict rotation and perspective transformation for lean postures (ablation) | GaitBase with QA and data augmentation · Gait3D# | Rank-1: 59.7 -> 61.1; 59.7 -> 77.2 (Rank-5) | Mixed |
| Use lean-aware rotation with random erasing and omit perspective transformation (ablation) | GaitBase with QA and data augmentation · Gait3D# | Rank-1: 59.7 -> 62.2 | Helped |
| Use sequence-level rather than frame-level alignment, with disturbance (ablation) | GaitBase with QA and data augmentation · Gait3D# | Rank-1: 63.0 -> 63.5 | Helped |
Surprising findings
- Quality assessment alone slightly lowers GaitBase on Gait3D (62.4% to 59.7% Rank-1) and GREW, which the authors attribute to mismatch with the original data augmentation.
Limitations
- Thresholds (epsilon=0.95, tau=0.001) were chosen via volunteer annotation and grid search rather than a principled criterion.
- Quality assessment alone can degrade some backbones (GaitBase on Gait3D and GREW), so gains depend on the pipeline and backbone.
- Quality indicator is a soft batch-normalized partial feature norm, without a validated ground-truth quality measure.
Experimental setup
Datasets: Gait3D GREW CASIA-B
Benchmarks: Gait3D GREW CASIA-B
Metrics: Rank-1 Rank-5 mAP R-10 R-20
Baselines: GaitSet GaitPart GLN GaitGL CSTL SMPLGait MTSGait DAGait GaitBase MetaGait DANet 3DLocal
Builds on: GaitBase ArcFace CosFace BNNeck Hu moments
Closest papers by meaning
- Exploring More from Multiple Gait Modalities for Human Identification (AAAI 2025 · similarity 0.88)
- On Denoising Walking Videos for Gait Recognition (CVPR 2025 · similarity 0.86)
- SkeletonGait: Gait Recognition Using Skeleton Maps (AAAI 2024 · similarity 0.85)
- RA-GAR: A Richly Annotated Benchmark for Gait Attribute Recognition (AAAI 2025 · similarity 0.85)
- BigGait: Learning Gait Representation You Want by Large Vision Models (CVPR 2024 · similarity 0.85)
- HybridGait: A Benchmark for Spatial-Temporal Cloth-Changing Gait Recognition with Hybrid Explorations (AAAI 2024 · similarity 0.84)
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.