Question about canonical camera parameters about metric3d HOT 2 CLOSED

Thanks for your great work. When I dive into the code about canonical ratio,I found that the canonical ratio was normalized by the real camera focal length x(f_x) in the part of preparing gt depth label(transform.py),but it was normalized by (f_x+f_y)/2 in the test script(do_test.py).Could the inconsistency lead to deep confusion? Looking forward to your reply!

In 【transform.py】, the transformation is used for training, where we resize the image while keeping the aspect ratio. The resize factor is defined according to the ratio of x-axis intrinsic.
However during inference in 【do_test.py】, we cannot garantee that the aspect ratio is kept for an unknown image. Hence we use an average ratio to de-transform.

from metric3d.

Thanks for your reply!

from metric3d.