Open suzdl opened 8 months ago
solve it by the following code
#mde_array = zoe.infer_pil(rgb_pil, output_type="numpy")*1000
mde_array = zoe.infer_pil(rgb_pil, pad_input=False, output_type="numpy")*1000
the parameter pad_input may make a confusing image with padded border in NYUv2, like this
this confusing operation not only affects the resolution of RGB input, but also affects the focal length of NYUv2.
I load the model "ZoeD_M12_N.pt", and run the general following code with little modification , to save metrics depth map first.
in this way, I will get the output metric depth in mm, and I test it by the following code, got bad rmse of 0.375 by the following code:
I compute the rmse in the same way for the method "NeWCRFs", and got the right rmse of 0.333. However, when I tried zoedepth, I got the rmse of 0.375. As you can see in my annotation, I have tried to crop the black border of RGB the same as Zoedepth, changed the mode of 'inter' into 'eval', and also tried to save the npy file to avoid round-off error, but none of them work. An interesting thing is that when I directly run the evaluate.py by the author, I got the right rmse of 0.27.
Directly utilize "mde_array = zoe.infer_pil(rgb_pil, output_type="numpy")" is so convenient to save depth map, but what makes the metrics down?