Multi-view contrastive learning for unsupervised 3D model retrieval and classification

Abstract

Unsupervised 3D model retrieval and classification have attracted a lot of attention due to wide applications. Although much progress has been achieved, they remain challenging due to the lack of supervised information to optimize neural network learning. Existing unsupervised methods usually utilized clustering algorithms to generate pseudo labels for 3D models. However, the clustering algorithms cannot fully mine the multi-view structure information and misguide the unsupervised learning process due to the noise information. To cope with the above limitation, this paper proposes a Multi-View Contrastive Learning (MVCL) method, which fully takes advantage of multi-view structure information to optimize the neural network. Specifically, we propose a multi-view grouping scheme and multi-view contrastive learning scheme to mine the self-supervised information and learn discriminative feature representation. The multi-view grouping scheme divides the multiple views of each 3D model into two groups and minimizes the group-level difference, which facilitates exploring the internal characteristics of 3D structural information. To learn the relationships among multiple views in an unsupervised manner, we propose a two-stream asymmetrical framework including the main network and the subsidiary network to guarantee the discrimination of the learned feature. Extensive 3D model retrieval and classification experiments are conducted on two challenging datasets, demonstrating the superiority of this method.