Learning-Deep-Learning

GeoNet: Unsupervised Learning of Dense Depth, Optical Flow and Camera Pose

July 2020

tl;dr: Use ResFlowNet and consistency check to improve monodepth.

Overall impression

GeoNet decouples the pixel movement to rigid flow and object motion adaptively. The movement of static parts in a video is solely caused by camera motion. The movement of dynamic object is caused by camera motion and specific object motion.

However it is still using photometric error instead of geometric error (unlike in DeepV2D).

Key ideas

Technical details

Notes