Articulation in Prime: Primitive-Based Articulated Object Understanding from a Single Casual Video

Researchers introduced a category-agnostic framework for 3D articulated object understanding from a single casual video.

The method fits geometric primitives and joint types to jointly recover part segmentation and kinematics, including revolute and prismatic motion.

It is designed to handle occlusions and fast camera ego-motion, making motion recovery more robust in real-world footage.

The approach is validated on new synthetic and real benchmarks, AiP-synth and AiP-real, and reduces reliance on category-specific training data.