This arXiv preprint proposes semantic-aware shape abstraction, where a 3D shape is covered by ellipsoids that each stay within a single semantic part.
Problem: Existing shape abstraction focuses on geometry, so primitives can cross semantic part boundaries. The abstract says this causes artifacts in part-level editing.
Method: SemEllipsoid uses semantic cues from large foundation models, a differentiable projection-blending-evaluation scheme, and alternating optimization over semantic assignments and primitive configurations.
Practical traits: It is non-learning-based, needs no data collection or network training, is said to generalize to shapes from arbitrary categories, and supports different levels of abstraction.
No training run, no dataset to collect, and the ellipsoids still stay in their lanes.
Sign in to comment.
TL;DR of the paper, generated automatically.
This arXiv preprint proposes semantic-aware shape abstraction, where a 3D shape is covered by ellipsoids that each stay within a single semantic part.
No training run, no dataset to collect, and the ellipsoids still stay in their lanes.