Train on own dataset with not good result about manydepth HOT 5 CLOSED

Hello, and thanks for trying out our repo! It's good to see that you are in fact able to learn some kind of depth map – it looks like you're in the right direction.

It looks from the images that you might perhaps have some problem with intrinsics. Did you update the intrinsics in the dataloader to match the intrinsics for your dataset?

It might be easier to try using the monodepth2 repo first (https://github.com/nianticlabs/monodepth2/) with your data, as monodepth2 is a simpler version of manydepth. This would help you with debugging. Then once you have good results there you should be able to copy your dataloaders across to manydepth, with minimal changes to make them work here.

I'll leave this issue open for a few more days in case anyone else has any other comments. Thanks

from manydepth.

Thanks a lot! I found out that I forgot the 25% crop in data preprocessing. I have corrected my intrinsic and dataloader problems and now I have more satisfactory results.

from manydepth.

Great news, these new results look great.

from manydepth.

hello,what does crop 25% mean？Thank you very much if you answer me

from manydepth.

Thanks a lot! I found out that I forgot the 25% crop in data preprocessing. I have corrected my intrinsic and dataloader problems and now I have more satisfactory results.

hello,what does crop 25% mean？Thank you very much if you answer me

from manydepth.