metric depth estimation finetuned with nuscenes

[xiaobh1519]

Hi, thanks for your great work!
I have tried a lot of experiments to fintune your metric depth estimation model with nuscenes dataset. But I find that the results turns worse ( “corrugations” appear) after about 3K steps like images below. The situation becomes more intense as training steps increases.



And I found that the effect of this phenomenon varies for images with different input resolutions. Can you tell me why this phenomenon occurs?
I have been troubled by this issue for a long time. I would be extremely grateful if I could get your answer.

[XYQ0121]

@xiaobh1519 Hello, I am also fine-tuning metric depth estimation using a custom dataset. I plan to use the KITTI config. My image size is 2160*3840. Can you tell me what modifications you made?Thanks a lot!

[XYQ0121]

Even though I changed the image size to 2160*3840 in the KITTI config, it does not match the printed values from the config. Is there any problem with this situation？

[LiheYoung]

Hi @xiaobh1519, did you try to decrease the learning rate of our pre-trained encoder?

[Won-JoonChoi]

I have an exactly same issue as @xiaobh1519 . Has anyone solved this problem?