search

nianticlabs / monodepth2

[ICCV 2019] Monocular depth estimation from a single image
Other
4.05k stars 949 forks source link

Negative loss #447

Closed nishavarghese15 closed 1 year ago

nishavarghese15 commented 1 year ago

Hi,

Thanks for your excellent work. I am trying to train the network using a custom dataset. Is there any chance of getting 'negative losses' during training? image

Thanks in advance!

zchoi commented 1 month ago

Have you solved this problem? I met the same issue.

nishavarghese15 commented 1 month ago

Have you solved this problem? I met the same issue.

Hi, I solved it. But I dont remember exactly what I did. But for a custom dataset, i have changed —max_depth, —min_depth options. Then I remember I had to use dataset with sufficient number of images. Initially I thought for a small number of images (around 200) it will overfit, but it didnt happen. I had to increase dataset size to thousands of images (>5000) for training. From my experience, this code will work for any dataset.

zchoi commented 1 month ago

Thanks