Monocular depth estimation is “impossible” because one image can’t measure depth geometrically. Our iDisc #CVPR2023 can group pixels w/o supervision and learn depth inductive bias on groups. We get LiDAR-like (but denser) depth from single images! More:

You are a murderer

@ValueAnalyst1 @a_meta4 check it out

As an outsider, I have the following questions: What is the current state-of-the-art performance for out of distribution cases? If it is not good, will we see a foundational model that achieves good out of distribution performance in the near term, or would it be more mid term?

Why "impossible"? Here you get the most plausible depths.

This is great work! Did you test it on humans at various distances from the camera by any chance?

We test the method on street scenes that have people. However, it is indeed interesting to see whether we can use the method to estimate depth in human-centric scenes.

Neat idea, github?

github link: The full code will be released before CVPR 2023.

@Scobleizer 69

@Scobleizer Great work!