Clarification on METRIC_DEPTH vs MODEL_METRIC_DEPTH usage in depth priors

[21yrm]

Hi, thanks for your fantastic work!

I have a small question regarding the metric estimation model attributes METRIC_DEPTH and MODEL_METRIC_DEPTH (as defined in vipe/vipe/priors/depth/base.py).

I'm not quite sure how to decide which one to use. Specifically, why do some models need to be re-run to update the depth estimation when the focal length changes, while others can simply adjust the output by multiplying with a new factor related to the focal length?

Could you please elaborate on how these two attributes are intended to be used, and in what scenarios one should be preferred over the other?

Thanks in advance for your clarification!

[21yrm]

I suppose one possible reason is that some models implicitly encode the camera intrinsics and use the encoded tokens to infer depth, while others treat the intrinsics as explicit factors that are not coupled with the depth inference itself, but rather used only for scale adjustment after the depth prediction.