3D Reconstruction from Single-View Image Using Feature Selection

Recovering the 3D shape of an object from single-view image with deep neural network has been attracting increasing attention in the past few years. Recent approaches based on convolutional neural networks have shown excellent results on single-view image. Most of them, however, have many model’s parameters or fewer parameters with performance degradation. Therefore, in this work we propose a feature selection module to balance this problem. This module first calculates the uncertain degree map to obtain the feature coordinates which means some coarse parts needs to be corrected. Then using these coordinates, features in several feature maps are selected. Finally, use MLP Layer to obtain fine features by taking features selected as input. Training and Inference are slightly different in this module. Using this module, we achieve better performance with about 18% parameters addition and comparable performance with about 30% model’s parameters decrease based on the Pix2Vox [1] framework.