3D Reconstruction from Images : A Review

Abstract

3D reconstruction from images is popular among the major branches of computer vision. The potential to acquire the exact 3D structure of an object or scene is still an issue for most of the important domains, like robot navigation, object recognition and scene understanding, 3D modeling and animation, industrial control and medical diagnosis. Different approaches of 3D reconstruction based on multi-view geometry and deep learning are reviewed. Several multi-view geometry based approaches including Structure-from-Motion (SfM), MultiView Stereo (MVS) and Surface Reconstruction are analysed. 3D reconstruction methods utilizing deep learning has attracted much attention recently. The High Resolution MVS Network (HighRes-MVSNet) and Three Dimentional Hierarchical Fusion Network (3D-FHNet) uses Convolutional Neural Network (CNN) for efficient 3D reconstruction. The efficient representation of 3D model is important for 3D reconstruction. Different 3D model representations and various training aspects for deep learning based approaches are summarized.

Keywords: 3D Reconstruction, Structure-from-Motion, Multi-View Stereo, Surface Reconstruction, HighRes-MVSNet, 3D-FHNet, Convolutional Neural Network