RingNet: Geometric Deep Representation Learning for 3D Multi-domain Protein Shape Retrieval

Abstract

AbstractMany applications use three-dimensional polygon meshes as geometric representation to achieve central tasks, such as 3D object retrieval and classification. However, implementing a deep learning approach dedicated to 3D meshes is a bit hard due to the complexity and irregularity of the mesh surface representation. In this paper, we propose a new geometric deep learning approach dedicated to representation learning, which applies convolutional operations on 3D meshes. In particular, we introduce a ring-unit convolutional operator that aggregates two graphs deduced from the mesh surface. Our network can learn highly discriminating features by avoiding complexity and irregularity problems.We experimentally validated our approach on 3D shape classification tasks and the multi-domain protein shape retrieval challenge. A comparison with the state-of-the-art approaches proved the relevance of the learned features to the accuracy of 3D object classification and retrieval.Keywords3D multi-domain protein shape retrievalGeometric deep learningDeep representation learningNeural network