Master’s thesis report

Link (.pdf): 3D Object Reconstruction with RGB, Depth and Normals Data

Abstract

Recently the field of computer vision has given rise to various machine learning based approaches to 3D object reconstruction. This project examines the state of the art progress in the field and improves upon one particular model. Neural Volumes by Facebook Reality Labs manages to reconstruct animated objects from RGB images taken from various view points. The inferred 3D model is photo realistic, but lacks the proper geometric shape of the ground truth object. We propose an extended model that, in addition to RGB images, uses depth and normal maps for supervision. Additionally, we introduce a script for Blender that allows us to render synthetic objects, on which to train the model. We evaluate the addition of depth and normals supervision quantitatively and qualitatively on several objects.

Some object reconstruction demos

Check all videos here!

Last modified on 2020-06-15