Skip to content

Latest commit

 

History

History
13 lines (7 loc) · 828 Bytes

edge_aware_depth_normal.md

File metadata and controls

13 lines (7 loc) · 828 Bytes

July 2019

tl;dr: Extend SfM-Learner by introducing a surface normal presentation. Precursor to LEGO.

Overall impression

The idea is good, that we introduce a surface normal map, which at each point should be perpendicular to the depth estimation.

However how to use it is a bit questionable. This work used normal map as an intermediate step (depth --> norm --> depth) and both conversion is deterministic by 3D geometry constraint. How this helps is puzzling. The result to be honest is not as good as claimed. You still see a lot of discontinuity of surface normals within the same object.

This work is superceded by their CVPR 2018 spotlight paper LEGO.