Tensor Multi-Task Learning for Multi-View Representation of 3D Shape

Abstract

Research on 3D representation has sprung up recently. However, most existing view-based methods lack interpretability and few researchers explore to mine high-order correlation for multi-view features aggregate in 3D shape representation. This paper presents a tensor multi-task model learning, T-MTL, which transforms the original multi-view classification problem into constructing a common space in which the multi-view representation can be compared. The proposed method can construct a tensor structure in the classifier weight subspace with multiple viewpoints and use a tensor low-rank constraint and Bregman discrepancy to capture high-order correlations between multi-view images and 3D models. In the classifier weight subspace, our method makes the inter-class divergence larger and the intra-class divergence smaller. Moreover, the proposed model can improve recognition performance further by unifying multiple visual features in a flexible manner. Abundant experiments on these challenging benchmark datasets show that our method achieves comparable results to the deep learning models approach.