Open innat opened 4 months ago
@innat, this is on our radar! Are you interested in adding this to KerasCV?
I think, ViT is required at first, so this PR https://github.com/keras-team/keras-cv/pull/1914 needs to be progressed as prior. Depth-anything model uses DINO-v2, which is not available in keras-cv.
Short Description
Depth Anything, a highly practical solution for robust monocular depth estimation by training on a combination of 1.5M labeled images and 62M+ unlabeled images.
Papers
https://depth-anything.github.io/
Existing Implementations
https://github.com/LiheYoung/Depth-Anything
Other Information
Prior requirement: dino-v2 | dinov1-keras - dinov2-keras
https://github.com/keras-team/keras-cv/issues/1628 https://github.com/keras-team/keras-cv/pull/1914 - this need to be progressed.