3D Semantic Shape Abstraction

越如 贾, Chengying Tu

Keywords: 3D Semantic Shape Abstraction

Abstract: The shape abstraction task is to learn the primitive-based representation of 3D objects. The past work usually learns semantic shape abstraction from manually labeled part annotation. Unsupervised learning for 3D semantic shape abstraction remains difficult because the model cannot understand the affordance of each part and the relationship between parts. In this project, we reproduce two off-the-shelf well-performing shape abstract models, Cuboid Shape Abstraction via Joint Segmentation and Neural Parts, on the PartNet dataset, and analyze their respective task perspectives and methodical characteristics. The former method uses the traditional cuboid primitive and tries to map the point cloud to a compact cuboid representation. And the latter proposes a new 3D primitive representation, which realizes homomorphic mapping between the sphere and the target object, and is more flexible than the traditional primitive representation. Our intuition is that Neural Parts will perform better on semantic shape abstraction. We test our intuition experimentally, but only in the chair category. Based on the experimental results, we analyze the advantages and disadvantages of the two.

TL;DR: We reproduce two 3D shape abstraction work and conduct qualitative and quantitative evaluations of their performance on PartNet.