Research map / 3D vision / Human pose estimation
Articulation in Prime: Primitive-Based Articulated Object Understanding from a Single Casual Video
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 a category-agnostic, per-video optimization that fits superquadric primitives to depth-backprojected point clouds to jointly recover part segmentation and revolute/prismatic joint parameters from a single casual monocular video, and introduces AiP-synth and AiP-real benchmarks.
Problem
Recovering part segmentation and joint kinematics (revolute/prismatic) of articulated objects from a single casually captured monocular video, where camera ego-motion, occlusions and weak local features break point-tracking methods, and learning-based methods fail to generalize across categories.
Approach
Treats articulated object understanding as primitive fitting: superquadric primitives act as a proxy representation, softly assigned to parts that carry screw-theory joint parameters, optimized jointly with a visibility-aware reconstruction loss and a scene-flow loss, with sparsity/existence regularizers that automatically prune superfluous parts and primitives.
Key results
- On AiP-real, mean mIoU 0.86 and axis error 6.53 degrees versus 0.38 mIoU and 41.38 degrees for Video2Articulation (best baseline), with 100% joint type accuracy.
- On Arti4D, revolute axis error 4.77 degrees and position error 5.44 cm versus Artipoint's 24.96 degrees and 85.49 cm.
- On AiP-synth prismatic joints, mIoU 0.94 and axis error 0.00 degrees versus 0.59 and 51.46 degrees for Video2Articulation; ReArt fails on all prismatic categories.
Surprising findings
- Ablating the reconstruction or flow losses sometimes yields joint type accuracy of 0% or 25%, showing that each loss term is individually critical.
Limitations
- Superquadric primitives give only coarse geometry, insufficient for detailed mesh reconstruction.
- Per-object optimization takes about 10 minutes on a single GPU and requires multiple seeds (5 per object).
- Relies on externally estimated depth, camera poses, scene flow (GMSF), and SAM2 masks; real-world AiP-real annotations are partly manual.
Experimental setup
Datasets: AiP-synth AiP-real Video2Articulation-S Arti4D PartNet-Mobility
Benchmarks: AiP-synth AiP-real Video2Articulation-S Arti4D
Metrics: mIoU axis error (deg) position error (cm) state error joint type accuracy
Baselines: Articulate-Anything Video2Articulation ReArt Artipoint
Builds on: superquadrics NeRF/3DGS alpha compositing GMSF ViPE SAM2
Closest papers by meaning
- Learning Implicit Representation for Reconstructing Articulated Objects (ICLR 2024 · similarity 0.90)
- Articulate your NeRF: Unsupervised articulated object modeling via conditional view synthesis (NeurIPS 2024 · similarity 0.88)
- Unsupervised Learning of Category-Level 3D Pose from Object-Centric Videos (CVPR 2024 · similarity 0.87)
- Towards Understanding Camera Motions in Any Video (NeurIPS 2025 · similarity 0.86)
- ArticulatedGS: Self-supervised Digital Twin Modeling of Articulated Objects using 3D Gaussian Splatting (CVPR 2025 · similarity 0.86)
- ImageNet3D: Towards General-Purpose Object-Level 3D Understanding (NeurIPS 2024 · 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.