Research map / 3D vision / Human pose estimation
R^2-Art: Category-Level Articulation Pose Estimation from Single RGB Image via Cascade Render Strategy
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.
Estimates per-part 6D articulation poses from a single RGB image by regressing root pose and depth, deforming a point-cloud prior via joint-centric kinematics, and refining with a three-stage cascade mask-rendering loss.
Problem
Category-level articulation pose estimation from a single RGB image, where prior methods require costly 3D point clouds or RGB-D input.
Approach
Joint-centric modeling that predicts per-joint states and applies Rodrigues-based kinematic transforms to a shape prior, combined with decoupled root pose/depth regression (projection depth times correction factor) and a coarse-to-fine cascade of area, box-alignment, and edge-aware pixel losses on the rendered mask.
Key results
- On ArtImage laptop, R2-Art (RGB only) reaches 37.5mm / 49.2mm ADD for the two parts, beating RGB-D baselines such as ContactArt (49.2mm / 62.1mm).
- Ablation on ArtImage drawer: full cascade strategy reduces ADD from 180.5mm to 105.6mm and raises AUC from 51.2% to 69.9%.
- On the real-world 7-part RobotArm dataset, R2-Art averages 46mm ADD and 68% AUC across parts, improving over the RGB-D A-NCSH baseline on each part.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use R2-Art instead of ContactArt | ArtImage Laptop pose estimation · ArtImage, Laptop | per-part ADD (mm): 49.2, 62.1 -> 37.5, 49.2 | Helped |
| Use R2-Art instead of ContactArt | ArtImage Eyeglasses pose estimation · ArtImage, Eyeglasses | per-part ADD (mm): 68.8, 118.1, 121.8 -> 63.4, 103.5, 106.5 | Helped |
| Use R2-Art instead of ContactArt | ArtImage Dishwasher pose estimation · ArtImage, Dishwasher | per-part ADD (mm): 95.3, 99.7 -> 83.5, 87.5 | Helped |
| Use R2-Art instead of ContactArt | ArtImage Scissors pose estimation · ArtImage, Scissors | per-part ADD (mm): 53.6, 57.8 -> 45.6, 48.3 | Helped |
| Use R2-Art instead of ContactArt | ArtImage Drawer pose estimation · ArtImage, Drawer | per-part ADD (mm): 81.3, 112.3, 145.8, 113.9 -> 72.3, 99.3, 126.3, 97.8 | Helped |
| Use R2-Art instead of A-NCSH | RobotArm pose estimation · RobotArm | per-part ADD (mm): 57.6, 62.2, 69.7, 75.5, 83.2, 89.4, 96.9 -> 37.6, 39.2, 45.3, 45.6, 48.9, 52.6, 60.8 | Helped |
| Use R2-Art instead of A-NCSH | RobotArm pose estimation · RobotArm | per-part AUC (%): 72.8, 69.2, 67.2, 63.4, 59.6, 58.3, 54.4 -> 80.2, 75.6, 72.3, 69.9, 64.2, 62.5, 59.2 | Helped |
| Add the complete cascade render strategy (ablation) | R2-Art · ArtImage, Drawer | ADD (mm): 180.5 -> 105.6 | Helped |
| Add the complete cascade render strategy (ablation) | R2-Art · ArtImage, Drawer | AUC (%): 51.2 -> 69.9 | Helped |
| Add area consistency loss (Stage I) (ablation) | R2-Art cascade render strategy · ArtImage, Drawer | ADD (mm): 180.5 -> 148.2 | Helped |
| Add alignment loss (Stage II) (ablation) | R2-Art cascade render strategy with area consistency · ArtImage, Drawer | ADD (mm): 148.2 -> 120.5 | Helped |
| Add pixel-wise refinement loss (Stage III) (ablation) | R2-Art cascade render strategy with area consistency and alignment · ArtImage, Drawer | ADD (mm): 120.5 -> 105.6 | Helped |
Surprising findings
- The RGB-only method outperforms several RGB-D baselines on ArtImage despite using no depth input.
Limitations
- Accumulated errors grow for deeper kinematic chains (parts 5-7 on RobotArm).
- Requires a category-level shape prior point cloud from the intra-category CAD models.
Experimental setup
Datasets: ArtImage ReArtMix RobotArm
Benchmarks: ArtImage ReArtMix RobotArm
Metrics: ADD (mm) AUC (%) 3D IoU (%)
Baselines: A-NCSH OMAD AKBNet ContactArt
Builds on: NOCS A-NCSH GPV-Pose
Closest papers by meaning
- SchemaPose: RGB-Based Category-Level Pose Estimation with Parametric Category Schema (NeurIPS 2026 · similarity 0.89)
- CAP-Net: A Unified Network for 6D Pose and Size Estimation of Categorical Articulated Parts from a Single RGB-D Image (CVPR 2025 · similarity 0.88)
- Unsupervised Learning of Category-Level 3D Pose from Object-Centric Videos (CVPR 2024 · similarity 0.87)
- GIVEPose: Gradual Intra-class Variation Elimination for RGB-based Category-Level Object Pose Estimation (CVPR 2025 · similarity 0.86)
- EfficientCAPER: An End-to-End Framework for Fast and Robust Category-Level Articulated Object Pose Estimation (NeurIPS 2024 · similarity 0.86)
- Real2Code: Reconstruct Articulated Objects via Code Generation (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.