Open JLeere opened 1 month ago
it can be an input, see https://github.com/apple/ml-depth-pro/issues/21#issuecomment-2397643379
it can be an input, see #21 (comment)
It can not. That is what my issue was about, the model does not take focal length as input. The code surrounding the model takes focal length as an argument, but that does not help the model make better predictions. It simply distorts the models output so that some large scale features get more accurate and some detailed features gets less accurate.
Do you think this issue can be corrected linearly calibrating with the specific setup to fit the scale?
I have one my own dataset , which means all images' focal length or camera matrix is already known. I want to know if this can be input to the model and improve the performance?