Learning probabilistic representation of shape recognition from volumetric grid

摘要

Since a camera only gets an observation of partial surface of an object, the recognition of the whole object from this single-view observation is imcomplete. To address this problem, we establish probability representation of surface shape recognition based on 3D data. For a number of objects, multiple Convolutional Neural Network (CNN) stuctures are used to extract local shape features and to fit probility representation. In experiments, computation effect of different network structures are compared and the feasibility of fitting probability representation using convolution neural network is verified. In most cases, recognition probability is very accurate.