Progressive Multimodal Shape Generation via Contextual Part Reasoning

Yixin Zhuang

Abstract: We present a contextual generative network for 3D shapes based on a conditional variational autoencoder, which learns a subspace of plausible complementary parts in the context of a partial shape. With the learned part subspace prior, which encodes bi-part spatial relations and geometry descriptions, a shape is generated via iterative “next part reasoning”, where a next part is sampled conditioned on a partial shape. Furthermore, our conditional subspace allows not just one, but a set of reasonable next parts to be generated, which adds controllability (e.g., via user selection) to the generative process. Our core idea of reasoning about next parts via conditional modeling offers a new way of understanding shape structures via part correlation modeling. Evaluations show the effectiveness of our approach and also the diversity of the generated shapes.