Research map / 3D vision / Human pose estimation
Instance-aware Contrastive Learning for Occluded Human Mesh Reconstruction
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 an instance-aware contrastive learning scheme that pulls joint features toward their owner's body-center feature and pushes apart centers of different people, improving occluded 3D human mesh reconstruction from a single image.
Problem
Monocular 3D human mesh reconstruction degrades under person-to-person occlusion because it is ambiguous which person a visible body part belongs to.
Approach
Encodes center-aligned and joint-aligned instance feature maps and trains them with a cosine-similarity contrastive loss (center-center push, center-joint pull) so that features carry per-person identity at body-center and joint locations; the contrastive part is used only at training time.
Key results
- On 3DPW-PC, the proposed method (trained without CrowdPose) reports MPJPE 99.9 and PA-MPJPE 76.3 versus 119.7 and 79.7 for ROMP*.
- On OCHuman, AP 34.9 and AP50 80.0 for the full model, compared with 24.8 AP for OCHMR*.
- On CMU-Panoptic, average MPJPE of 126.1 is the best among compared methods, versus 127.6 for 3DCrowdNet.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add instance-aware contrastive learning with center and joint features (ablation) | Ours · 3DPW-PC | MPJPE: 104.2 -> 102.0 | Helped |
| Add instance-aware contrastive learning with center and joint features (ablation) | Ours · 3DPW-PC | PA-MPJPE: 80.2 -> 77.2 | Helped |
| Add positional encoding to the center-aligned instance map (ablation) | Ours · 3DPW-PC | MPJPE: 102.2 -> 102.0 | Helped |
| Add positional encoding to the center-aligned instance map (ablation) | Ours · 3DPW-PC | PA-MPJPE: 78.2 -> 77.2 | Helped |
| Set center and joint feature dimension to 32 instead of 64 (ablation) | Ours · 3DPW-PC | MPJPE: 102.0 -> 103.5 | Hurt |
| Set center and joint feature dimension to 128 instead of 64 (ablation) | Ours · 3DPW-PC | MPJPE: 102.0 -> 103.1 | Hurt |
| Set center and joint feature dimension to 256 instead of 64 (ablation) | Ours · 3DPW-PC | MPJPE: 102.0 -> 104.2 | Hurt |
| Use the proposed method instead of ROMP | 3D human mesh reconstruction · 3DPW-PC | MPJPE: 115.6 -> 102.0 | Helped |
| Use the proposed method instead of OCHMR* | 3D human mesh reconstruction · OCHuman | AP: 24.8 -> 34.9 | Helped |
| Use the proposed method instead of ROMP | 3D human mesh reconstruction · CrowdPose | AP50: 58.8 -> 65.7 | Helped |
| Use the proposed method instead of 3DCrowdNet | 3D human mesh reconstruction · CMU-Panoptic | MPJPE: 127.6 -> 126.1 | Helped |
| Use the proposed method instead of CLIFF | 3D human mesh reconstruction · 3DPW | MPJPE: 72.0 -> 78.0 | Hurt |
Surprising findings
- On the full 3DPW test set, the ResNet-50 version (78.0 MPJPE) is worse than several prior methods, and CoordFormer and others outperform it there, which the authors attribute to occlusion-focused design rather than general performance.
Limitations
- Gains on the general 3DPW benchmark are not evident; ResNet-50 variant trails several methods on full 3DPW.
- Contrastive loss uses ground-truth center and joint positions for feature sampling during training.
Experimental setup
Datasets: Human3.6M MPI-INF-3DHP MuCo-3DHP MPII LSP COCO CrowdPose
Benchmarks: 3DPW-PC OCHuman CrowdPose CMU-Panoptic 3DPW
Metrics: MPJPE PA-MPJPE MPVPE AP AP50 AP75 APM APL
Baselines: SPIN PyMAF ROMP CLIFF HMDiff CoordFormer OCHMR 3DCrowdNet CRMH BMP PARE METRO
Builds on: ROMP OCHMR 3DCrowdNet SMPL
Closest papers by meaning
- DPMesh: Exploiting Diffusion Prior for Occluded Human Mesh Recovery (CVPR 2024 · similarity 0.87)
- Self-Supervised 3D Human Mesh Recovery from a Single Image with Uncertainty-Aware Learning (AAAI 2024 · similarity 0.86)
- FactorizedHMR: A Hybrid Framework for Video Human Mesh Recovery (NeurIPS 2026 · similarity 0.86)
- BLADE: Single-view Body Mesh Estimation through Accurate Depth Estimation (CVPR 2025 · similarity 0.85)
- ODHSR: Online Dense 3D Reconstruction of Humans and Scenes from Monocular Videos (CVPR 2025 · similarity 0.84)
- AiOS: All-in-One-Stage Expressive Human Pose and Shape Estimation (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.