July 2019
tl;dr: Estimate the intrinsics in addition to the extrinsics of the camera from any video.
Same authors for Struct2Depth. This work eliminates the assumption of the availability of intrinsics. This opens up a whole lot possibilities to learn from a wide range of videos.
This network regresses depth, ego-motion, object motion and camera intrinsics from mono videos. Thus it is named learn-K (intrinsics) --> The idea of regressing intrinsics is similar to GLNet.
- Estimate each of the intrinsics. It learns four parameters, 2 for optical centers (principal points) and 2 for focal length.
- Occlusion aware loss (picking the most foreground pixels during photometric loss calculation)
- Foreground mask to mask out the possible moving objects.
- Use a randomized layer optimization (this is quite weird)
- Sometimes an overall supervision signal is given to two tightly coupled parameters and it is not enough to get accurate estimate for both parameters. (cf. Deep3Dbox)
- In detail, how was the lens correction regressed?
- See interview with the CEO of isee on this paper.
- Q: Can we project the intermediate representation (3D points) to BEV instead of back to camera plane for loss calculation? This would eliminate the need for using occlusion-aware loss.
- code is based on that of struct2depth.