Research map / 3D vision / Face and shape
SAOR: Single-View Articulated Object 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.
Learns single-view 3D shape, texture, and viewpoint of articulated animals with a skeleton-free part-based linear blend skinning model and a balanced-sampling swap loss, trained only on silhouettes and relative depth from pretrained networks.
Problem
Single-view 3D reconstruction of articulated object categories (e.g., animals) without category-specific 3D templates, skeletons, or multi-view training data.
Approach
A skeleton-free part-based articulation model (predicted soft part assignments with per-part scale/rotation/translation LBS) combined with a simplified cross-instance swap loss using random pairs and silhouette-clustered balanced sampling to enforce viewpoint diversity and avoid degenerate flat solutions.
Key results
- On CUB, SAOR-Bird achieves 51.9 PCK (57.8 w/o aquatic), best among methods without 3D templates, skeletons, or extra data.
- On PASCAL quadrupeds, SAOR-101 reaches 44.9 PCK on horse, 33.6 on cow, 29.1 on sheep, outperforming CSM, A-CSM and Unicorn, and beating MagicPony on cow and sheep but not horse (42.9 vs 44.9 on horse for SAOR).
- Ablation shows removing the swap loss drops horse PCK from 44.9 to 30.8, and removing articulation drops it to 26.3.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| remove depth supervision (ablation) | SAOR-101 · Pascal horse, cow, and sheep; CUB bird | PCK@0.1: 42.4, 30.1, 26.8, 49.9 -> 44.9, 33.6, 29.1, 51.9 | Helped |
| add the swap loss (ablation) | SAOR · Pascal horse, cow, and sheep; CUB bird | PCK@0.1: 30.8, 17.7, 18.4, 44.5 -> 44.9, 33.6, 29.1, 51.9 | Helped |
| add balanced viewpoint sampling (ablation) | SAOR · Pascal horse, cow, and sheep; CUB bird | PCK@0.1: 27.5, 20.1, 18.3, 38.8 -> 44.9, 33.6, 29.1, 51.9 | Helped |
| add articulation (ablation) | SAOR · Pascal horse, cow, and sheep; CUB bird | PCK@0.1: 26.3, 19.4, 17.9, 41.7 -> 44.9, 33.6, 29.1, 51.9 | Helped |
| use SAOR-101 instead of MagicPony | Pascal horse · Pascal horse | PCK@0.1: 42.9 -> 44.9 | Helped |
| use SAOR-101 instead of MagicPony | Pascal cow · Pascal cow | PCK@0.1: 42.5 -> 33.6 | Hurt |
| use SAOR-101 instead of MagicPony | Pascal sheep · Pascal sheep | PCK@0.1: 26.2 -> 29.1 | Helped |
| use SAOR-101 instead of MagicPony | Animal3D · Animal3D horse and cow | normalized L1 Chamfer error: 0.046 -> 0.046; 0.040 -> 0.043 | Mixed |
Surprising findings
- Swap loss without articulation or balanced sampling produces flat degenerate shapes for quadrupeds, though it works for birds.
Limitations
- Texture predictions are not fully realistic.
- Fails on unusual viewpoints far from training distribution or partially visible objects.
- Depends on estimated silhouettes and relative depth from pretrained networks.
Experimental setup
Datasets: CUB LSUN iNaturalist PASCAL VOC Animal3D SAM masks MiDaS depth
Benchmarks: CUB PASCAL quadrupeds Animal3D
Metrics: PCK@0.1 keypoint transfer normalized L1 Chamfer distance
Baselines: CMR U-CMR DOVE MagicPony UMR Unicorn CSM A-CSM Dense-Equi
Builds on: Unicorn CMR MagicPony SMR pose transfer skeleton-free representation
Closest papers by meaning
- MagicPony: Learning Articulated 3D Animals in the Wild (CVPR 2023 · similarity 0.90)
- LASR: Learning Articulated Shape Reconstruction From a Monocular Video (CVPR 2021 · similarity 0.90)
- BANMo: Building Animatable 3D Neural Models From Many Casual Videos (CVPR 2022 · similarity 0.89)
- S3O: A Dual-Phase Approach for Reconstructing Dynamic Shape and Skeleton of Articulated Objects from Single Monocular Video (ICML 2024 · similarity 0.87)
- REACTO: Reconstructing Articulated Objects from a Single Video (CVPR 2024 · similarity 0.86)
- SINGAPO: Single Image Controlled Generation of Articulated Parts in Objects (ICLR 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.