Shape recognition with recurrent neural network

Abstract

Shape recognition is a fundamental problem in the field of computer vision and is important to various applications. A number of methods based on deep CNN has acquired state-of-the-art performance in shape recognition. Among them, model-based methods perform convolutions with 3D filters on the voxels or point cloud in continuous 3D space, and the volumetric representation makes them exploit complete structure information. Unfortunately, in order to train the deep network with available samples in a reasonable amount of time, these methods have to use a coarse representation, typically 30×30×30 grid, which will inevitably sacrifice much discriminate detail information. This paper presents a novel approached based on recurrent neural network to solve this problem. In each step, the model selects the location of the subvolume from where the local 3D CNN feature is extracted, and the hypothesis is formulated by merging the features of subvolumes of each step. In this way, the proposed approach can explore the shape with high resolutions and exploit the fine-grained 3D structure information. Primary experimental results on the public dataset verify the effectiveness of the proposed method.