Unsupervised Monocular Depth Learning in Dynamic Scenes

November 2020

tl;dr: Learning depth and motion without explicit semantic masks.

Overall impression

Challenges in SfM: textureless areas, occlusions, reflections, and moving objects. Most previous work deal with dynamic scene by semantic segmentation and ignore pixels in them. Monodepth2 proposed a method to detect pixels do not change across frames to exclude these regions in training.

This paper is the extension of LearnK to dynamic scenes. This paper proposed a sparsity regularization on motion field and it seems to work quite well. It learns four parameters, 2 for optical centers (principal points) and 2 for focal length.

Key ideas

Technical details