Machine Learning for Image Based 3D Reconstruction

Accurate 3D reconstruction (MVS) often encounters significant challenges when dealing with textureless surfaces.

In this master thesis, we assume that textureless areas are smooth and piecewise planar, we present a novel approach by leveraging plane segmentation (CNN) to improve to improve the accuracy and completeness of depth and normal map estimation, particularly in scenes where traditional photo-consistency-based methods fall short.

Finally; we evaluate our algorithm in terms of accuracy and completeness on the ETH-3D high resolution dataset.

Keywords: Multi-View Stereo, Plane Detection and Segmentation, Convolutional Neural Network.