Learning-Deep-Learning

FSM: Full Surround Monodepth from Multiple Cameras

April 2021

tl;dr: 360 deg monodepth prediction from self-supervision.

Overall impression

This paper pushes the application and the frontier of monodepth. It extends the scope of the previous works on monodepth from single camera to a multi-camera setting.

FSM is intended for multi-camera config with very large baselines (and thus minimal image overlap). Stereo rectification and disparity estimation is not feasible.

Along the line of view synthesis, FSM introduces two new constraints for optimization and one benefit: STC (spatio-temporal constraint) and PCC (pose consistency constraints), and scale-awareness introduced by metric scale extrinsics.

The FSM assumes known camera extrinsics (more precisely, it only requires relative extrinsics between cameras).

Key ideas

Technical details

Notes