Learning-Deep-Learning

Monodepth2: Digging Into Self-Supervised Monocular Depth Estimation

June 2019

tl;dr: Three tricks to handle occlusion/dis-occlusion/relatively static objects/static scenes better.

Overall impression

Monodepth is on estimating depth using stereo pairs. This paper digs into videos. It is heavily influenced by the CVPR2017 paper sfm learner. The changes to losses are surprisingly simple, yet effective. In particular, the selecting of the min of many losses is very similar to the idea in FSAF.

This paper has a very solid code base which has become the foundation of many future works (such as D3VO).

Key ideas

Technical details

Notes