Research map / 3D vision / Human pose estimation
Continuous Heatmap Regression for Pose Estimation via Implicit Neural Representation
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 NerPE, an implicit-neural-representation decoder that regresses continuous heatmaps for pose estimation, enabling arbitrary-resolution heatmap output and progressive coordinate decoding.
Problem
Discrete heatmap representations in human pose estimation introduce quantization errors that degrade accuracy, especially at low input resolutions where heatmap-based methods fall behind coordinate regression.
Approach
Generalize sub-pixel convolution's upsampling factor to infinity to obtain a continuous upsampling function; an MLP decoder takes a local feature vector and a relative query coordinate and is trained with uniformly sampled queried positions to fit Gaussian/Laplace ground-truth heatmaps, so heatmap resolution is decoupled from input resolution. A progressive coarse-to-fine decoding avoids generatin
Key results
- On COCO val at 64x64 input with ResNet-50, NerPE reaches 40.8 AP vs 34.4 for SimpleBaseline and 39.3 for SimCC.
- On COCO test-dev with HRNet-W48 at 384x288, NerPE achieves 76.2 AP, versus 76.0 for SimCC and 75.5 for HRNet.
- On CrowdPose at 64x64 with HRNet-W32, NerPE improves AP from 42.4 (HRNet) to 47.4.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace discrete heatmap regression with continuous heatmap regression | ResNet-50 pose estimation · COCO val, 64 × 64 input | AP: 39.3 -> 40.8 | Helped |
| replace discrete heatmap regression with continuous heatmap regression | HRNet-W48 pose estimation · COCO val, 64 × 64 input | AP: 59.7 -> 62.5 | Helped |
| replace discrete heatmap regression with continuous heatmap regression | TokenPose-S pose estimation · COCO val, 64 × 64 input | AP: 62.8 -> 64.4 | Helped |
| replace discrete heatmap regression with continuous heatmap regression | HRNet-W48 pose estimation · COCO test-dev | AP: 76.0 -> 76.2 | Helped |
| replace discrete heatmap regression with continuous heatmap regression | HRNet-W32 pose estimation · MPII, 128 × 128 input | PCKh@0.1: 32.6 -> 34.6 | Helped |
| replace discrete heatmap regression with continuous heatmap regression | HRNet-W32 pose estimation · CrowdPose, 64 × 64 input | AP: 46.5 -> 47.4 | Helped |
| replace discrete heatmap regression with continuous heatmap regression | HRNet-W32 pose estimation · CrowdPose, 256 × 192 input | AP: 66.7 -> 66.9 | Helped |
| add uniform position sampling (ablation) | NerPE with ResNet-50 · MPII, 128 × 128 input, 4 × 4 cell division | PCKh@0.5: 80.83 -> 81.42 | Helped |
| add uniform position sampling (ablation) | NerPE with ResNet-50 · MPII, 128 × 128 input, 8 × 8 cell division | PCKh@0.5: 82.54 -> 82.65 | Helped |
| use 8 × 8 rather than 2 × 2 cell division (ablation) | NerPE with uniform sampling · MPII, 128 × 128 input | PCKh@0.5: 76.03 -> 82.65 | Helped |
| increase samples per cell from 4 to 64 (ablation) | NerPE without uniform sampling · MPII, 128 × 128 input, 8 × 8 cells | PCKh@0.5: 80.98 -> 82.54 | Helped |
| increase samples per cell from 4 to 64 (ablation) | NerPE with uniform sampling · MPII, 128 × 128 input, 8 × 8 cells | PCKh@0.5: 81.45 -> 82.65 | Helped |
Limitations
- INR decoder is deliberately simple; the network structure is not optimized for INR characteristics.
- No error bars or statistical significance reported.
- Gains at standard high resolution are small (e.g., +0.2 AP at 256x192).
Experimental setup
Datasets: COCO MPII CrowdPose
Benchmarks: COCO MPII CrowdPose
Metrics: AP AR PCKh@0.5 PCKh@0.1 GFLOPS
Baselines: SimpleBaseline HRNet SimCC DARK TokenPose
Builds on: sub-pixel convolution local implicit image function (LIIF) SimpleBaseline HRNet DARK TokenPose
Closest papers by meaning
- Neural Localizer Fields for Continuous 3D Human Pose and Shape Estimation (NeurIPS 2024 · similarity 0.85)
- PoseGen: Learning to Generate 3D Human Pose Dataset with NeRF (AAAI 2024 · similarity 0.85)
- Video-Based Human Pose Regression via Decoupled Space-Time Aggregation (CVPR 2024 · similarity 0.84)
- Attention-Propagation Network for Egocentric Heatmap to 3D Pose Lifting (CVPR 2024 · similarity 0.84)
- MotionMap: Representing Multimodality in Human Pose Forecasting (CVPR 2025 · similarity 0.83)
- NRDF: Neural Riemannian Distance Fields for Learning Articulated Pose Priors (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.