Research map / 3D vision / Human pose estimation
DynPose: Largely Improving the Efficiency of Human Pose Estimation by a Simple Dynamic Framework
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 DynPose, a dynamic top-down pose estimation framework where a lightweight router scores per-person difficulty and routes samples to either a small or large pretrained HPE model.
Problem
Top-down human pose estimation is inefficient because pose samples are highly diverse and detectors produce many redundant boxes, so a single large model wastes computation on easy or meaningless samples.
Approach
A lightweight router is trained to predict a difficulty score from the image, supervised by the normalized OKD gap between a small and a large pretrained network; at inference, only mid-difficulty ("Hard") samples go to the large network, with thresholds set by a target split ratio, and the router transfers across models and datasets without retraining.
Key results
- Dyn-R50-H32/50 reaches 74.8 AP on COCO val vs 74.9 for HRNet-W32 while running at 270 FPS vs 223 FPS.
- Dyn-R152-VL/50 reaches 78.0 AP vs 78.2 for ViTPose-L, increasing FPS from 59 to 93 and reducing FLOPs from 58.68G to 35.90G.
- Router trained on ResNet-50/HRNet-W32 transfers without fine-tuning to other pairs, e.g., ResNet-50 with ViTPose-L keeps 78.0 AP at 98 FPS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use router-controlled routing instead of random routing (ablation) | DynPose with ResNet-50 and HRNet-W32 · COCO | AP | Helped |
| Use DynPose routing with ResNet-50 and HRNet-W32 at a 50% split | ResNet-50 and HRNet-W32 · COCO val | FPS: 223 -> 270 | Helped |
| Use DynPose routing with ResNet-152 and ViTPose-L at a 50% split | ResNet-152 and ViTPose-L · COCO val | AP: 78.2 -> 78.0 | No clear effect |
| Use DynPose routing with ResNet-152 and ViTPose-L at a 50% split | ResNet-152 and ViTPose-L · COCO val | FPS: 59 -> 93 | Helped |
| Transfer the COCO-trained router to CrowdPose and route 60% of samples to HRNet-W32 | Dyn-R50-H32 · CrowdPose | AP: 67.5 -> 67.3 | No clear effect |
| Add image pooling to the router (ablation) | DynPose router · Router ablation | inference time (ms): 0.52 -> 0.10 | Helped |
| Add PConv to the pooled router (ablation) | DynPose router · Router ablation | AP: 74.71 -> 74.73 | Helped |
| Add SGE to the pooled router (ablation) | DynPose router · Router ablation | AP: 74.71 -> 74.78 | Helped |
| Use the full router with pooling, PConv, and SGE rather than pooling alone (ablation) | DynPose router · Router ablation | AP: 74.71 -> 74.81 | Helped |
| Set θ to 0.1 rather than other tested values (ablation) | DynPose inference strategy · COCO val | AP: 74.78; 74.81; 74.68; 74.48; 73.91 | Helped |
Surprising findings
- Roughly 140K detected boxes for only about 6K real samples on COCO val, with about half of routed samples falling into an 'Extremely Hard' band caused by redundant or incorrect boxes.
- Random routing to the large network is worse than router-controlled routing at all split ratios, but the gain is measured against a simple baseline only.
Limitations
- Transfer to CrowdPose required a larger split ratio (60%) and a different theta (0.3), so the thresholds do not transfer cleanly across distributions.
- Speed measured on a single NVIDIA 3090 GPU; the router adds overhead and results depend on detector-produced box sets.
Experimental setup
Datasets: COCO 2017 keypoint detection
Benchmarks: COCO val 2017 COCO test-dev CrowdPose
Metrics: AP AP50 FPS FLOPs memory usage
Baselines: HRNet-W32 ViTPose-L ResNet-50 ResNet-152 Swin-L TokenPose PPT OKDHP DistilPose SDPose Lite-HRNet
Builds on: HRNet ViTPose OKS SGE DynamicDet
Closest papers by meaning
- Dynamic Inertial Poser (DynaIP): Part-Based Motion Dynamics Learning for Enhanced Human Pose Estimation with Sparse Inertial Sensors (CVPR 2024 · similarity 0.84)
- Dynamic Camera Poses and Where to Find Them (CVPR 2025 · similarity 0.84)
- ProbPose: A Probabilistic Approach to 2D Human Pose Estimation (CVPR 2025 · similarity 0.84)
- ManiPose: Manifold-Constrained Multi-Hypothesis 3D Human Pose Estimation (NeurIPS 2024 · similarity 0.83)
- Sequential Joint Dependency Aware Human Pose Estimation with State Space Model (AAAI 2025 · similarity 0.83)
- Dynamic Support Information Mining for Category-Agnostic Pose Estimation (CVPR 2024 · 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.