Research map / 3D vision / Human pose estimation
You Only Learn One Query: Learning Unified Human Query for Single-Stage Multi-Person Multi-Task Human-Centric Perception
Sheng Jin, Shuhuai Li, Tong Li, Wentao Liu, Chen Qian, Ping Luo
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| co-train pose with detection and segmentation (add P to D+S) (ablation) | HQNet R-50 · COCO-UniHuman val detection | AP: 73.0 -> 74.5 | Improved |
| co-train pose with detection and segmentation (add P to D+S) (ablation) | HQNet R-50 · COCO-UniHuman val segmentation | AP: 63.6 -> 65.7 | Improved |
| add classification (gender and age) tasks to D+S+P (ablation) | HQNet R-50 · COCO-UniHuman val detection | AP: 74.5 -> 74.9 | Improved |
| add classification tasks to D+S+P (ablation) | HQNet R-50 · COCO-UniHuman val keypoints | AP: 69.5 -> 69.3 | No clear effect |
| add Gender-aided human Model Selection (GaMS) for SMPL model choice (ablation) | HQNet R-50 · COCO-UniHuman val 3D mesh | PA-MPJPE (lower better): 54.92 -> 50.80 | Improved |
| add GaMS (ablation) | HQNet R-50 · COCO-UniHuman val 3D mesh | MPJPE (lower better): 87.00 -> 84.74 | Improved |
| use HQ-Ins matching (detection, pose, segmentation losses) instead of detection-only matching (ablation) | HQNet · qualitative | matching errors | Improved |
| train queries with multiple tasks instead of detection only (ablation) | HQNet R-50 queries as re-ID features · PoseTrack21 MOT | IDF1: 62.4 -> 64.6 | Improved |
| add auxiliary denoising losses for segmentation and pose (ablation) | HQNet · COCO-UniHuman | overall | No clear effect |
| use ViT-L instead of R-50 backbone | HQNet · COCO-UniHuman val keypoints | AP: 69.3 -> 75.3 | Improved |
| single-stage unified model versus top-down ViTPose with ViT-L | 2D keypoint estimation · COCO-UniHuman val | keypoint AP: 78.2 vs 75.3 | Hurt |
| freeze backbone and fine-tune HQNet for new face detection task | HQNet R-50 · COCO-UniHuman val face detection | AP: 58.2 -> 68.4 | Improved |
Closest papers by meaning
- Human-MME: A Holistic Evaluation Benchmark for Human-Centric Multimodal Large Language Models (ICLR 2026 · similarity 0.85)
- L-Man: A Large Multi-modal Model Unifying Human-centric Tasks (AAAI 2025 · similarity 0.84)
- Divide and Conquer: Hybrid Pre-training for Person Search (AAAI 2024 · similarity 0.84)
- HumanPCR: Probing MLLM Capabilities in Diverse Human-Centric Scenes (ICLR 2026 · similarity 0.84)
- A Large-Scale Human-Centric Benchmark for Referring Expression Comprehension in the LMM Era (NeurIPS 2024 · similarity 0.84)
- Sharingan: A Transformer Architecture for Multi-Person Gaze Following (CVPR 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.
