ArtModeling: An Instance-Matched Dataset and Benchmark for Articulated-Object Perception and Dynamics

Anonymous Authors

01 / THE DATA

Captured Object Views

Section 1: Start and end states, supplemental views, and object masks from real-world captures.

01

Articulation States

Cabinet 16 captured at its start state
START
Cabinet 16 captured at its end state
END
02

Supplemental Views

Cabinet 16 photographed from a supplemental pose to reveal its underside
ADDITIONAL POSE
03

Object Masks

Cabinet 16 isolated on a black background
MASKED RGB

02 / INSTANCE EXPLORER

Dataset Showcase

Section 2: Explore each object's captures, interactions, USD asset motion, and physical properties.

03 / INTERACTION

RGB and Depth Interactions

Section 3: Aligned RGB and depth sequences show real-world articulated motion.

01 / RGB CAPTURE
Interaction sequence of a drawer cabinet opening
cabinet16
02 / DEPTH CAPTURE
Colorized depth sequence of the same drawer cabinet interaction
cabinet16

Physical Fidelity Evaluation

Section 4: We use an optical–inertial motion capture system to record the joint trajectories of real-world objects. We compare these recordings with the joint motion of their corresponding articulated digital twins to evaluate differences between simulated and real-world joint behavior.

Asset Comparison with Astra

Section 5: Compare ArtModeling assets with Astra reconstructions from 30 and 240 images.

SolverAdapt

Section 6: Inspect USD joints, preview motion, and edit physical properties with the bundled local Python service.

Explore joint drives, motion limits, mass, friction, and gravity.

Download sample USD models ↗
Open SolverAdapt