what kind of depth image is expected to have best performance with current checkpoint?

[lyc233333]

thanks for your great jobs in this project!

i am trying to use it to refine the depth image got from rgbd/stereo cameras, but it works not so great. could you please tell something more about the depth image used in this project? like the format, range, precision etc, and i can add more preprocessing on it.

[TychoBomer]

Searching for the same thing!

[charisoudis]

Hi, I am also exploring its application to RGBD data. It seems that the model does not generalize well to different prompt depth domains, in contrast with DepthAnything v2 (metric checkpoint). In particular, the fusion modules are expecting low res depth maps to draw sparse scale-shifts from and perform (sprarse) local refinement of the mono depths. When prompted with higher-res depth maps, whose values contain larger local variances, fusion performs worse.