Please advise on the issue of inaccurate depth maps

in = np.array([[[1252.81310, 0.00000000, 826.588115],
[0.00000000, 1252.81310, 469.984663],
[0.00000000, 0.00000000, 1.00000000]]])
print("000",in.shape)

images = sorted(glob.glob("/root/autodl-tmp/val-night/"+processed_line))
prediction = model.inference(
images,
intrinsics=in,
align_to_input_ext_scale=True )
This is a very high-quality work, but we are experiencing some issues with inaccurate depth maps.

I set `align_to_input_ext_scale=True`, but the output depth map is (1, 280, 504), while the actual input image dimensions are (3, 900, 1600). I also obtained a poor d1 value（after resize pred）.

I believe this is a very good work, and the problem might be due to my incorrect operation. If you have time, please help me. Thank you very much.

No repository file or test is named. Start by reproducing the shown model.inference call with the supplied intrinsics and image dimensions, then inspect how align_to_input_ext_scale affects the returned depth shape and how the resized prediction is evaluated. Done means determining whether the output dimensions and poor d1 value come from usage or an implementation issue.

Written by the indexing model from the issue text.