Learning-Deep-Learning

Manydepth: The Temporal Opportunist: Self-Supervised Multi-Frame Monocular Depth

January 2022

tl;dr: Efficient and accurate use of multiframe information for monodepth.

Overall impression

For monodepth application, sequence information is often available at test time. There exists two ways to leverage multiframes for monodepth estimation. First uses expensive test-time refinement techniques (Consistent Video Depth, Robust CVD, CoMoDa, SSIA, etc) or recurrent network (). ManyDepth is an adaptive approach to dense depth estimation that can make use of seq info at test time when it is available.

The paper provides a good overview of recent advances of self-supervised monodepth.

ManyDepth address what was thought to be a forced choice in 3D reconstruction, between classic triangulation over multiple frames versus instant-but-fragile single-frame inference with a neural network. (Source)

Key ideas

Technical details

Notes