How to train a model without background?

Totoro97 / NeuS

Code release for NeuS

MIT License

1.52k stars 205 forks source link

How to train a model without background? #53

Open ThePrecoder opened 2 years ago

ThePrecoder commented 2 years ago

I have the photos like this: rgba000 Shall I use the config womask.conf?And then whether the Nerf++ parameter should be added without background.Thanks a lot!

Totoro97 commented 2 years ago

Hi, you may use the wmask.conf and the the parameter train.use_white_bkgd as True.

ThePrecoder commented 2 years ago

Thanks for your reply.Another question is that have you ever tried to run this code in outdoor scenes?I have the pictures like 000000 I use the pretrained segmentation model to mark out the sky part and give that area a weight loss to lower the weight sum to 0.However,the result just not so good.And I dont know how to do it.

Totoro97 commented 2 years ago

Hi I guess the segmented images might not be very multi-view consistent (i.e., a same background is masked out in image A while not masked out in image B), which could cause some inconsistency across views and may mislead the reconstruction.

ThePrecoder commented 2 years ago

Hi I guess the segmented images might not be very multi-view consistent (i.e., a same background is masked out in image A while not masked out in image B), which could cause some inconsistency across views and may mislead the reconstruction.

Okay,thanks

nangongliufeng commented 1 year ago

Hi I guess the segmented images might not be very multi-view consistent (i.e., a same background is masked out in image A while not masked out in image B), which could cause some inconsistency across views and may mislead the reconstruction.

Okay,thanks

Have you solved your problem? I have a similar issue. Can you please explain the solution in detail?