3D shape estimation in a constraint optimization neural network

Abstract

One of the most important aspects of visual perception is the inference of 3D shape from a 2D retinal image of the outside world. The existence of several valid mapping functions from object to data makes this inverse problem ill-posed and therefore computationally difficult. In human vision, the retinal image is a 2D projection of the 3D world. The visual system imposes certain constraints on the family of solutions in order to uniquely and efficiently solve this inverse problem. This work specifically focused on the minimization of standard deviations of 3D angles (MSDA) for 3D perception. Our goal was to use a Deep Convolutional Neural Network based on biological principles derived from visual area V4 to achieve 3D reconstruction using constrained minimization of MSDA. We conducted an experiment using novel shapes with human subjects to collect data and test the model. The performance of the network largely agreed with how humans estimated novel 3D shapes. The results show that the constraint of MSDA in 3D shape can be implemented in a neural network and produce human-like results. Additional visual constraints can be added to the network in the future to fully test the theory of visual constraints as a basis of 3D shape perception.