Research map / 3D vision / Human pose estimation
KPA-Tracker: Towards Robust and Real-Time Category-Level Articulated Object 6D Pose Tracking
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 KPA-Tracker, a category-level articulated object 6D pose tracker that registers unsupervised, ordered 3D keypoints (learned by KPA-Gen) between canonicalized adjacent frames, with keyframe-based tracking for robustness and real-time speed.
Problem
Category-level online 6D pose tracking of articulated objects from point cloud or RGB-D video, where dense per-pixel methods are slow and sensitive to self-occlusion.
Approach
Formulate articulated pose tracking as per-part 3D keypoint registration: canonicalize the input with the previous frame's pose, learn ordered category-level keypoints without human annotation (KPA-Gen with chamfer, separation and coverage losses), predict inter-frame pose increments, and refine via keyframe-based tracking.
Key results
- On PartNet-Mobility Laptop, KPA-Tracker reaches 5.0/7.8 degree rotation error and 0.05 s inference time, versus CAPTRA's 5.9/5.3 degrees and 0.10 s.
- Tracking speed of 14.2 FPS, versus 10.7 FPS for CAPTRA, 3.3 for OMAD and 0.8 for A-NCSH.
- On the semi-synthetic ReArt-48 Box category, per-part rotation errors are 5.6 and 5.9 degrees with translation errors of about 0.01 m.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use 8 keypoints per part instead of 4 (ablation) | KPA-Tracker · PartNet-Mobility Laptop | rotation error | Mixed |
| use 8 keypoints per part instead of 4 (ablation) | KPA-Tracker · PartNet-Mobility Laptop | translation error | Mixed |
| use 12 keypoints per part instead of 4 (ablation) | KPA-Tracker · PartNet-Mobility Laptop | rotation error | Hurt |
| use 12 keypoints per part instead of 4 (ablation) | KPA-Tracker · PartNet-Mobility Laptop | translation error | Hurt |
| use 16 keypoints per part instead of 4 (ablation) | KPA-Tracker · PartNet-Mobility Laptop | rotation error | Mixed |
| use 16 keypoints per part instead of 4 (ablation) | KPA-Tracker · PartNet-Mobility Laptop | translation error | Hurt |
| use KPA-Tracker instead of CAPTRA* | articulated pose tracking · PartNet-Mobility | inference time: 0.10 -> 0.05; 0.14 -> 0.08; 0.11 -> 0.06; 0.12 -> 0.06; 0.25 -> 0.12 | Helped |
| use KPA-Tracker instead of CAPTRA | articulated pose tracking · PartNet-Mobility | FPS: 10.7 -> 14.2 | Helped |
Surprising findings
- Increasing the number of keypoints beyond 4 per part degraded tracking performance, with 16 keypoints giving worse rotation error on laptops than 4.
Limitations
- Real-world RobotArm results show notable errors (rotation up to 20 degrees, translation up to 0.16 m), attributed to multi-depth structure.
- Performance on some categories (e.g., Drawer, Eyeglasses) is not uniformly better than CAPTRA; gains over CAPTRA in rotation are slight.
Experimental setup
Datasets: PartNet-Mobility ReArt-48 RobotArm
Benchmarks: PartNet-Mobility synthetic tracking ReArt-48 semi-synthetic tracking RobotArm
Metrics: rotation error (degrees) translation error (m) inference time FPS
Baselines: A-NCSH OMAD Oracle ICP CAPTRA
Builds on: CAPTRA PointNet++ 6-PACK ICP
Closest papers by meaning
- 6D Object Pose Tracking in Internet Videos for Robotic Manipulation (ICLR 2025 · similarity 0.87)
- ProxyPose: 6-DoF Pose Tracking via Video-to-Video Translation (NeurIPS 2026 · similarity 0.85)
- R^2-Art: Category-Level Articulation Pose Estimation from Single RGB Image via Cascade Render Strategy (AAAI 2025 · similarity 0.85)
- ComPose: When to Trust Hands for Object Pose Tracking (NeurIPS 2026 · similarity 0.85)
- EfficientCAPER: An End-to-End Framework for Fast and Robust Category-Level Articulated Object Pose Estimation (NeurIPS 2024 · similarity 0.85)
- Instance-Adaptive and Geometric-Aware Keypoint Learning for Category-Level 6D Object Pose Estimation (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.