Could you please provide the specific environment configuration?

[daydreamer2023]

Hi!
Thank you for your generous sharing！

I can't get an acceptable result on KITTI, neither depth or motion.
Have you ever obtained results comparable to those in the paper？

[chamorajg]

I have not trained it on KITTI dataset, but I have applied the same procedure on a different private dataset where I got good results. I will train it and check it on KITTI dataset and release the model. The GPU I own is too small to perform training on that. Maybe if you could help me like narrow down the mistake where it is happening, I shall be happy to help or fix it.

[hkjournalist01]

Hello @chamorajg! I also can't get acceptable results on KITTI, the depth inference outputs a blank image (almost same value every pixel). I noticed that the official paper initialized the encoder part of depth network from a network pretrained on ImageNet, is this initialization missing in your implementation?