Depth Detection Based on Multi-Scale Residual Multi-Layer Perceptron Iterative Network

Abstract

Three-dimensional (3D) reconstruction can be used to satisfy the requirement of people on 3D cinema, 3D games, 3D medical imaging and 3D map. The traditional method on 3D reconstruction needs large computation and possesses low accuracy. With the development of AI, more attention to convolutional neural network for three-dimensional reconstruction is put. Current 3D reconstruction methods limit the robustness and integrity and reduce the accuracy of reconstructed models when dealing with occluded, texture-free, or low-textured local backgrounds. Aiming at the problems of low accuracy and cost volume construction in the current 3D reconstruction process, a 3D reconstruction method based on iterative refinement of multi-scale residuals was proposed. The basic idea is to obtain the initial depth map by feature extraction to generate point clouds, and then use multi-scale feature fusion pyramid multi-layer perceptron (MLP) network to obtain point features at different levels. This new multi-scale residual MLP iterative network is used to predict the depth value, and the residual between the depth prediction and the real value is used to estimate the loss. This method attains more precise depth information that enhances the ability of 3D reconstruction of objects, simplifies the network model, and reduces the computational burden. Investigational outcomes show that the suggested method can be used in processing the binocular vision with occluded, texture-free or low-textured local backgrounds and generating higher quality 3D reconstructed objects than the previous methods.