Research map / 3D vision / Human pose estimation
BEDLAM2.0: Synthetic humans and cameras in motion
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.
Introduces BEDLAM2.0, a larger synthetic video dataset with diverse focal lengths, synthetic and captured camera motions, richer body shapes, hair, shoes, and clothing, showing that training HPS/world-coordinate regressors on it improves accuracy over BEDLAM.
Problem
Estimating 3D human motion in world coordinates from video with moving cameras is hard because existing synthetic training data lacks ground-truth human and camera motion together with diverse focal lengths and realistic appearance.
Approach
A synthetic dataset rendered in Unreal Engine 5.3 that adds diverse focal lengths, zoom, many synthetic camera motion types with shake, real handheld and Apple Vision Pro camera trajectories, graded clothing, strand-based hair, shoes, and more body shapes and 3D environments than BEDLAM.
Key results
- CameraHMR trained on BEDLAM2 reduces 3DPW PA-MPJPE from 43.2 (BEDLAM1) to 41.1 and MPJPE from 68.0 to 64.8.
- On EMDB world-space, PromptHMR trained on B1+B2 reaches WA-MPJPE100 of 70.5 vs 77.6 for B1, and 71.0 for the original PromptHMR.
- Training on BEDLAM2 reduces PVE-T-SC shape error by 18.6% relative to BEDLAM1 (8.85 to 7.20).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train CameraHMR on BEDLAM2 instead of BEDLAM1 | CameraHMR · 3DPW | MPJPE: 68.0 -> 64.8 | Helped |
| Train CameraHMR on BEDLAM2 instead of BEDLAM1 | CameraHMR · EMDB | MPJPE: 88.7 -> 74.6 | Helped |
| Train CameraHMR on BEDLAM2 instead of BEDLAM1 | CameraHMR · RICH | MPJPE: 75.2 -> 70.8 | Helped |
| Train CameraHMR on BEDLAM2 instead of BEDLAM1 | CameraHMR · B2 test set | PVE-T-SC: 8.85 -> 7.20 | Helped |
| Train GVHMR on BEDLAM2 instead of BEDLAM1 | GVHMR · RICH | WA-MPJPE₁₀₀: 87.3 -> 75.5 | Helped |
| Train GVHMR on BEDLAM2 instead of BEDLAM1 | GVHMR · EMDB | WA-MPJPE₁₀₀: 112.4 -> 113.7 | Hurt |
| Train PromptHMR on BEDLAM2 instead of BEDLAM1 | PromptHMR · RICH | WA-MPJPE₁₀₀: 85.7 -> 75.3 | Helped |
| Train PromptHMR on BEDLAM2 instead of BEDLAM1 | PromptHMR · EMDB | WA-MPJPE₁₀₀: 77.6 -> 71.9 | Helped |
| Train CameraHMR on BEDLAM1+2 instead of BEDLAM1 | CameraHMR · 3DPW | MPJPE: 68.0 -> 65.2 | Helped |
| Train CameraHMR on BEDLAM1+2 instead of BEDLAM2 | CameraHMR · EMDB | MPJPE: 74.6 -> 75.5 | Mixed |
| Train GVHMR on BEDLAM1+B2 instead of BEDLAM1 | GVHMR · RICH | WA-MPJPE₁₀₀: 87.3 -> 75.8 | Helped |
| Train PromptHMR on BEDLAM1+B2 instead of BEDLAM1 | PromptHMR · RICH | WA-MPJPE₁₀₀: 85.7 -> 72.5 | Helped |
Surprising findings
- Training on BEDLAM2 alone gives a lower shape error than the combined B1+B2 setting for CameraHMR (7.20 vs 7.44 PVE-T-SC).
- Removing sitting and stair-climbing motions from BEDLAM2 reduces accuracy on stair climbing in EMDB relative to BEDLAM1.
Limitations
- Only supports people interacting with the ground, not with objects or other humans; motions are not semantically meaningful in the scene.
- Visual domain gap remains between synthetic and real video; no facial motion or audio.
- Population coverage excludes children, amputees, and people with physical impairments.
Experimental setup
Datasets: BEDLAM BEDLAM2.0 AMASS MOYO BEAT2 ARCTIC Google Scanned Objects CAESAR PolyHaven HDRI Flickr HumanFOV
Benchmarks: 3DPW EMDB RICH BEDLAM2.0 test set
Metrics: MPJPE PA-MPJPE PVE WA-MPJPE100 W-MPJPE100 RTE Jitter Foot-Sliding PVE-T-SC
Baselines: CameraHMR GVHMR PromptHMR DPVO+HMR2.0 GLAMR TRACE SLAHMR WHAM
Builds on: BEDLAM SMPL-X AMASS CameraHMR GVHMR PromptHMR
Closest papers by meaning
- HumanVid: Demystifying Training Data for Camera-controllable Human Image Animation (NeurIPS 2024 · similarity 0.89)
- WHAM: Reconstructing World-grounded Humans with Accurate 3D Motion (CVPR 2024 · similarity 0.87)
- Lifting Motion to the 3D World via 2D Diffusion (CVPR 2025 · similarity 0.87)
- MVHumanNet: A Large-scale Dataset of Multi-view Daily Dressing Human Captures (CVPR 2024 · similarity 0.87)
- ODHSR: Online Dense 3D Reconstruction of Humans and Scenes from Monocular Videos (CVPR 2025 · similarity 0.87)
- Reconstructing People, Places, and Cameras (CVPR 2025 · similarity 0.86)
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.