Research map / Detection and segmentation / CNN features
PBADet: A One-Stage Anchor-Free Approach for Part-Body Association
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 PBADet, a one-stage anchor-free detector that adds a single 2D part-to-body center offset to anchor-free YOLO-style heads to associate human parts with their bodies.
Problem
Detecting human parts (hands, face, feet) and associating each with the correct person is typically done with multi-stage pipelines, anchor-based heads, or body-to-part offsets that scale poorly with the number of parts and break when parts are occluded.
Approach
Extend anchor-free per-pixel detection with one 2D part-to-body center offset, supervised with TOOD-style task-aligned anchor selection, so each part maps one-to-one to a body regardless of the number of parts.
Key results
- On BodyHands, PBADet (YOLOv7, 36.9M params) reaches 89.1 hand AP, 92.62 conditional accuracy and 85.98 joint AP, above BPJDet (YOLOv5l6) at 85.9 / 86.91 / 84.39.
- On COCOHumanParts, PBADet (YOLOv5l6) achieves 43.0 AP.5:.95 (41.8 subordinate) vs BPJDet (YOLOv5l6) 43.6 (43.3), with lowest average mMR^-2 on head, face, and hands among compared methods.
- Versus ED-Pose on COCO val for hands and feet, PBADet reports all-parts F1 of 48.52 vs 31.51.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the part-body association loss and prediction head (ablation) | PBADet with YOLOv7 · BodyHands | Joint AP: 78.07 -> 85.98 | Helped |
| use multi-scale anchor-relative center offsets instead of normalized absolute distances (ablation) | PBADet with YOLOv7 · BodyHands | Joint AP: 85.46 -> 85.98 | Helped |
| use task alignment rather than all positive samples for association supervision (ablation) | PBADet with YOLOv7 · BodyHands | Joint AP: 85.78 -> 85.98 | Helped |
| use anchor-free rather than anchor-based detection (ablation) | PBADet part-to-body with YOLOv7 · BodyHands | Joint AP: 85.28 -> 85.98 | Helped |
| use part-to-body rather than body-to-part association (ablation) | anchor-based YOLOv7 · BodyHands | Joint AP: 78.13 -> 85.28 | Helped |
Surprising findings
- On COCOHumanParts, the overall AP.5:.95 of PBADet is slightly below BPJDet (YOLOv5l6) (43.0 vs 43.6), so gains are not uniform across metrics.
Limitations
- Gains over BPJDet are small on COCOHumanParts and CrowdHuman face-body association.
- Single-part face-body task shows less advantage over BPJDet, as acknowledged by the authors.
Experimental setup
Datasets: BodyHands COCOHumanParts CrowdHuman AnimalPose AP-10K
Benchmarks: BodyHands COCOHumanParts CrowdHuman AP-10K
Metrics: AP conditional accuracy joint AP MR^-2 mMR^-2 precision recall F1
Baselines: OpenPose Keypoint Com. MaskRCNN+FD MaskRCNN+FS MaskRCNN+LD MaskRCNN+IoU BodyHands BPJDet (YOLOv5s6) BPJDet (YOLOv5m6) BPJDet (YOLOv5l6) Hier R-CNN ED-Pose
Builds on: YOLOv7 YOLOv5 TOOD BPJDet FCOS
Closest papers by meaning
- PairDETR : Joint Detection and Association of Human Bodies and Faces (CVPR 2024 · similarity 0.87)
- Detangling People: Individuating Multiple Close People and Their Body Parts via Region Assembly (CVPR 2017 · similarity 0.85)
- Realtime Multi-Person 2D Pose Estimation Using Part Affinity Fields (CVPR 2017 · similarity 0.85)
- GigaHumanDet: Exploring Full-Body Detection on Gigapixel-Level Images (AAAI 2024 · similarity 0.84)
- Learning from Synchronization: Self-Supervised Uncalibrated Multi-View Person Association in Challenging Scenes (CVPR 2025 · similarity 0.83)
- PoseBH: Prototypical Multi-Dataset Training Beyond Human Pose Estimation (CVPR 2025 · similarity 0.83)
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.