Research map / 3D vision / Human pose estimation
BLADE: Single-view Body Mesh Estimation through Accurate Depth Estimation
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.
BLADE estimates the pelvis depth Tz from a single image, conditions a ControlNet-style AiOS pose estimator on Tz, and solves focal length and XY-translation by differentiable mask alignment, plus a 2M-image close-range synthetic dataset BEDLAM-CC.
Problem
Single-image human mesh recovery with accurate perspective projection: jointly estimating body shape, pose, focal length and 3D translation fails for close-range people with strong perspective distortion, where orthographic/heuristic camera models cause 3D pose and 2D alignment errors.
Approach
Perspective distortion is driven by the person's Z-translation Tz (not focal length), so Tz is estimated first from the image with a depth backbone, used to condition the mesh estimator, and then focal length and XY-translation are solved by differentiable rasterization against the person mask, avoiding orthographic-to-perspective heuristics.
Key results
- On SPEC-MTP, BLADE reduces Tz error by 85.9% relative to ZOLLY and PVE by 21.4%.
- On BEDLAM-CC, BLADE improves mIoU by 44.8% over ZOLLY (74.6 vs 51.8 mIoU reported in Table 1 for Ours vs ZOLLY).
- Pelvis depth backbone ablation on SPEC-MTP: DAv2 gives Tz error 15.4 cm; final BLADE version reaches 12.7 cm.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use BLADE (real-world training) instead of ZOLLY | BLADE · SPEC-MTP | ETz: 0.899 -> 0.127 | Helped |
| Use BLADE (real-world training) instead of ZOLLY | BLADE · SPEC-MTP | mIoU: 62.3 -> 69.5 | Helped |
| Use BLADE (real-world training) instead of ZOLLY | BLADE · BEDLAM-CC | mIoU: 51.8 -> 75.0 | Helped |
| Use DiNOv2 instead of DAv2 as depth-estimator backbone (ablation) | BLADE pelvis depth estimator · SPEC-MTP | ETz: 0.154 -> 0.300 | Hurt |
| Use Sapiens instead of DAv2 as depth-estimator backbone (ablation) | BLADE pelvis depth estimator · SPEC-MTP | ETz: 0.154 -> 0.210 | Hurt |
| Add improved augmentation and BEDLAM-CC data to the DAv2-based depth estimator (ablation) | BLADE pelvis depth estimator · SPEC-MTP | ETz: 0.154 -> 0.127 | Helped |
| Fine-tune AiOS on HUMMAN, PDHUMAN, and H36M without Tz conditioning (ablation) | BLADE pose estimator · SPEC-MTP | PVE: 110.851 -> 120.582 | Hurt |
| Add Tz conditioning through the ControlNet-style architecture (ablation) | AiOS pose estimator · SPEC-MTP | PVE: 110.851 -> 99.635 | Helped |
Surprising findings
- Good performance on the synthetic PDHUMAN dataset does not translate to real-world close-range performance; SOTA methods perform well on SPEC-MTP but poorly on PDHUMAN in PVE.
- Naively fine-tuning AiOS on close-range datasets degrades generalization to real-world SPEC-MTP images (PVE 110.9 to 120.6).
Limitations
- Only single-person images are handled.
- Lens distortion and non-pinhole cameras such as fisheye are not modeled.
- Focal length and translation estimation can fail when the segmentation mask is inaccurate.
Experimental setup
Datasets: Human3.6M PDHUMAN HUMMAN BEDLAM-CC SPEC-MTP 3DPW BEDLAM
Benchmarks: SPEC-MTP PDHUMAN BEDLAM-CC HUMMAN 3DPW Human3.6M
Metrics: PVE mIoU focal length error Tz error XY translation error inverse Tz error PA-MPJPE MPJPE
Baselines: ZOLLY AiOS TokenHMR SMPLer-X SPEC
Builds on: AiOS Depth Anything V2 ControlNet ZOLLY BEDLAM
Closest papers by meaning
- Self-Supervised 3D Human Mesh Recovery from a Single Image with Uncertainty-Aware Learning (AAAI 2024 · similarity 0.88)
- TokenHMR: Advancing Human Mesh Recovery with a Tokenized Pose Representation (CVPR 2024 · similarity 0.88)
- FactorizedHMR: A Hybrid Framework for Video Human Mesh Recovery (NeurIPS 2026 · similarity 0.87)
- Reconstructing Humans with a Biomechanically Accurate Skeleton (CVPR 2025 · similarity 0.87)
- Reconstructing People, Places, and Cameras (CVPR 2025 · similarity 0.87)
- AiOS: All-in-One-Stage Expressive Human Pose and Shape Estimation (CVPR 2024 · similarity 0.87)
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.