Research map / 3D vision / Human pose estimation
DualPM: Dual Posed-Canonical Point Maps for 3D Shape and Pose 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.
Introduces Dual Point Maps (DualPMs), a pair of canonical and posed point maps with a layered amodal extension, trained purely on synthetic renders of one or two quadruped models to reconstruct 3D shape and pose from a single image.
Problem
Monocular 3D shape and pose reconstruction of deformable quadruped objects, where a single point map gives geometry but not the deformation relative to a canonical rest pose, and visible-only maps miss self-occluded parts.
Approach
Predict two pixel-aligned point maps from one image: a posed camera-space point map and a canonical-space point map (viewpoint- and pose-invariant), so the deformation field is their difference; extend to layered amodal point maps via depth-peeling-style layers, and condition posed prediction on the canonical map.
Key results
- On PASCAL VOC keypoint transfer, DualPM reaches PCK@0.1 of 73.2% on horses versus 53.9% for 3D-Fauna.
- On Animodel, DualPM achieves real-sized Chamfer distance of 3.13 cm on horses versus 5.11 cm for Trellis and 8.69 cm for 3D-Fauna.
- Ablation: predicting bone rotations (supervised 3D-Fauna-style) yields 4.96 cm Chamfer on horses vs 3.13 cm for the DualPM model; conditioning P on Q alone outperforms conditioning on F and Q (3.13 vs 3.29 cm).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DualPM reconstruction instead of 3D-Fauna | Quadruped reconstruction · PASCAL VOC Horse | PCK@0.1 (%): 53.9 -> 73.2 | Helped |
| Use DualPM reconstruction instead of MagicPony | Quadruped reconstruction · PASCAL VOC Cow | PCK@0.1 (%): 42.5 -> 66.86 | Helped |
| Use DualPM reconstruction instead of MagicPony | Quadruped reconstruction · PASCAL VOC Sheep | PCK@0.1 (%): 41.2 -> 66.82 | Helped |
| Use DualPM reconstruction instead of Trellis | Quadruped reconstruction · Animodel Horse | Chamfer Distance (cm), real-sized: 2.23 ± 1.39 -> 1.52 ± 0.88 | Helped |
| Use DualPM reconstruction instead of Trellis | Quadruped reconstruction · Animodel Cow | Chamfer Distance (cm), real-sized: 1.85 ± 1.31 -> 1.11 ± 0.73 | Helped |
| Use DualPM reconstruction instead of Trellis | Quadruped reconstruction · Animodel Sheep | Chamfer Distance (cm), real-sized: 1.96 ± 1.22 -> 1.33 ± 0.78 | Helped |
| Predict point maps instead of bone rotations (ablation) | DualPM on Animodel Horses · Animodel Horses | Chamfer Distance (cm), real-sized: 4.96 ± 1.66 -> 3.13 ± 2.13 | Helped |
| Predict point maps instead of bone rotations (ablation) | DualPM on Animodel Horses · Animodel Horses | Chamfer Distance (cm), normalized: 1.78 ± 0.59 -> 1.11 ± 0.73 | Helped |
| Use 8 layers instead of 4 (ablation) | DualPM on Animodel Horses · Animodel Horses | Chamfer Distance (cm), real-sized: 3.13 ± 2.13 -> 3.13 ± 1.98 | No clear effect |
| Use 8 layers instead of 4 (ablation) | DualPM on Animodel Horses · Animodel Horses | Chamfer Distance (cm), normalized: 1.11 ± 0.73 -> 1.11 ± 0.69 | No clear effect |
| Condition P on image features F and canonical point maps Q rather than Q alone (ablation) | DualPM on Animodel Horses · Animodel Horses | Chamfer Distance (cm), real-sized: 3.13 ± 2.13 -> 3.29 ± 2.26 | Hurt |
| Condition P on image features F and canonical point maps Q rather than Q alone (ablation) | DualPM on Animodel Horses · Animodel Horses | Chamfer Distance (cm), normalized: 1.11 ± 0.73 -> 1.17 ± 0.80 | Hurt |
Surprising findings
- Conditioning the posed point map on the canonical map alone generalizes better out-of-distribution than also conditioning on image features.
- Trained on only one or two synthetic models per category, the method generalizes to real images and different breeds like foals and donkeys.
Limitations
- Evaluated mainly on quadrupeds (horses, cows, sheep); a separate model is trained per category.
- Layered amodal representation captures only a limited number of ray-surface intersections (4 vs 8 layers showed minimal difference).
Experimental setup
Datasets: Animodel PASCAL VOC Goat images from prior work
Benchmarks: PASCAL VOC keypoint transfer (PCK@0.1) Animodel bi-directional Chamfer distance
Metrics: PCK@0.1 bi-directional Chamfer distance (cm)
Baselines: A-CSM MagicPony Farm3D 3D-Fauna Trellis
Builds on: DUSt3R NOCS DINOv2 Farm3D 3D-Fauna
Closest papers by meaning
- DUSt3R: Geometric 3D Vision Made Easy (CVPR 2024 · similarity 0.86)
- Sparse-view Pose Estimation and Reconstruction via Analysis by Generative Synthesis (NeurIPS 2024 · similarity 0.86)
- SPAR3D: Stable Point-Aware Reconstruction of 3D Objects from Single Images (CVPR 2025 · similarity 0.86)
- MUSt3R: Multi-view Network for Stereo 3D Reconstruction (CVPR 2025 · similarity 0.85)
- Test3R: Learning to Reconstruct 3D at Test Time (NeurIPS 2025 · similarity 0.85)
- Spectral Meets Spatial: Harmonising 3D Shape Matching and Interpolation (CVPR 2024 · similarity 0.85)
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.