Multi-view Convolutional Neural Networks for 3D Shape Recognition

A longstanding question in computer vision concerns the representation of 3D objects for shape recognition: should 3D objects be represented with shape descriptors operating on their native 3D format, such as their voxel grid or polygon mesh, or can they be effectively represented with view-based descriptors? We address this question in the context of learning to recognize 3D objects from a collection of their rendered views on 2D images. We first present a standard CNN architecture trained to recognize the objects' rendered views independently of each other. Starting from such a network, we show that a 3D object can be recognized even from a single view at an accuracy far higher than using state-of-the-art 3D shape descriptors. The same architecture can be applied to accurately recognize human hand-drawn sketches of objects. Recognition rates further increase when multiple views of the objects are provided. In addition, we present a novel CNN architecture that combines information from multiple views of a 3D object into a single and compact shape descriptor offering even better recognition performance. We conclude that a collection of 2D views can be highly informative for 3D object recognition and is amenable to emerging CNN architectures and their derivatives.