Question about "Digging Into Self-Supervised Monocular Depth Estimation"

[jahaniam]

I was wondering if your network output for "self-supervised Stereo Training" is disparity or inverse depth?

"With known focal length and camera baseline the predicted disparity map, i.e. the inverse depth map, can then be converted into scaled metric depth"

Thanks

[mrharicot]

Hi,

The network outputs disp through a sigmoid, which is then turned into a depth through:

min_disp = 1 / max_depth
max_disp = 1 / min_depth
scaled_disp = min_disp + (max_disp - min_disp) * disp
depth = 1 / scaled_disp

Where we used max_depth = 100 and min_depth = 0.1

I hope this helps!

[daniyar-niantic]

@a-jahani The implementation is now public at https://github.com/nianticlabs/monodepth2