Research map / 3D vision / Human pose estimation
EfficientCAPER: An End-to-End Framework for Fast and Robust Category-Level Articulated Object Pose 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.
Proposes EfficientCAPER, an end-to-end two-stage network that estimates category-level articulated object poses by predicting the free part's pose with decoupled rotation and then constrained parts' poses via joint state/parameter prediction, without post-hoc optimization.
Problem
Category-level 6D pose estimation for articulated objects suffers from ignoring kinematic constraints, self-occlusion, and the need for time-consuming optimization or retrieval post-processing.
Approach
Joint-centric pose modeling: the free part's pose is regressed directly with decoupled rotation (two confidence-weighted axis normals), then the input is canonicalized by that pose and constrained parts' poses are recovered from predicted joint parameters and joint states via the Rodrigues formula (revolute) or translation (prismatic), replacing intermediate NOCS-style representations and optimiza
Key results
- On ArtImage, EfficientCAPER reports 2.5-3.1 deg rotation error on Dishwasher and 0.02 s inference per image versus 0.8-16.5 s for baselines (A-NCSH, OMAD, ArtPERL).
- On ReArtMix, rotation error improves over ReArtNet across all categories (e.g., Box 0.9/1.7 deg vs 3.4/3.6 deg).
- Self-occlusion analysis on Drawer shows rotation error stays around 1.9-2.0 deg across 0%-100% occlusion levels; translation error rises to 0.095 m at 80-100% occlusion.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add articulation pose canonicalization (ablation) | EfficientCAPER · ArtImage Laptop | joint state error: 3. 5◦ -> 1.8◦ | Helped |
| add articulation pose canonicalization (ablation) | EfficientCAPER · ArtImage Eyeglasses | joint state error: 7.8◦, 8.3◦ -> 5.0◦, 3.7◦ | Helped |
| add articulation pose canonicalization (ablation) | EfficientCAPER · ArtImage Scissors | 3D scale score (%): 57.2, 56.0 -> 56.5, 49.7 | Mixed |
| add articulation pose canonicalization (ablation) | EfficientCAPER · ArtImage Drawer | 3D scale score (%): 58.9, 57.5, 64.7, 65.2 -> 71.3, 59.4, 66.5, 68.4 | Helped |
| input free part and constrained parts instead of only the free part (ablation) | EfficientCAPER stage one · ArtImage Dishwasher | rotation error (°): 20.4 -> 2.5 | Helped |
| input free part and constrained parts instead of only the free part (ablation) | EfficientCAPER stage one · ArtImage Eyeglasses | translation error (m): 0.053 -> 0.037 | Helped |
| input free part and constrained parts instead of only the free part (ablation) | EfficientCAPER stage one · ArtImage Scissors | rotation error (°): 5.0 -> 5.0 | No clear effect |
| input free part and constrained parts instead of only the free part (ablation) | EfficientCAPER stage one · ArtImage Laptop | translation error (m): 0.059 -> 0.061 | Hurt |
| input free part and constrained parts instead of only the free part (ablation) | EfficientCAPER stage one · ArtImage Drawer | rotation error (°): 3.4 -> 2.0 | Helped |
| input free part and constrained parts instead of only the free part (ablation) | EfficientCAPER stage one · ArtImage Laptop | rotation error (°): 3.8 -> 3.5 | Helped |
| use joint-centric pose modeling | EfficientCAPER · ArtImage Drawer, increasing occlusion | rotation error (°): 1.9, 2.0, 2.0 | No clear effect |
| use EfficientCAPER instead of ReArtNet | ReArtMix · ReArtMix Box | rotation error (°): 3.4, 3.6 -> 0.9, 1.7 | Helped |
Surprising findings
- Some categories show worse results than baselines, e.g., Scissors rotation error 5.0/5.1 deg on ArtImage is higher than A-NCSH and ArtPERL, and Eyeglasses translation errors are sometimes higher than baselines.
Limitations
- Relies heavily on free-part pose estimation accuracy.
- Small datasets and relatively simple structures limit the demonstrated superiority.
- Evaluated only on static pose estimation; trained per category separately.
Experimental setup
Datasets: ArtImage ReArtMix RobotArm
Benchmarks: ArtImage ReArtMix RobotArm
Metrics: rotation error (deg) translation error (m) 3D IoU (%) inference time (s)
Baselines: A-NCSH OMAD ArtPERL ReArtNet
Builds on: GPV-Pose HS-Pose A-NCSH OMAD NOCS
Closest papers by meaning
- R^2-Art: Category-Level Articulation Pose Estimation from Single RGB Image via Cascade Render Strategy (AAAI 2025 · similarity 0.86)
- Unsupervised Learning of Category-Level 3D Pose from Object-Centric Videos (CVPR 2024 · similarity 0.86)
- Articulation in Prime: Primitive-Based Articulated Object Understanding from a Single Casual Video (NeurIPS 2026 · similarity 0.86)
- GenCOPE: Syn2Real Generalized Category-Level Object Pose Estimation for Robotic Picking (NeurIPS 2026 · similarity 0.85)
- KPA-Tracker: Towards Robust and Real-Time Category-Level Articulated Object 6D Pose Tracking (AAAI 2024 · similarity 0.85)
- Vision Foundation Model Enables Generalizable Object Pose Estimation (NeurIPS 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.