search

JonasSchult / Mask3D

Mask3D predicts accurate 3D semantic instances achieving state-of-the-art on ScanNet, ScanNet200, S3DIS and STPLS3D.
MIT License
525 stars 103 forks source link

An error when training S3DIS. #134

Open LinLin1031 opened 11 months ago

LinLin1031 commented 11 months ago

I changed my custom dataset to have the exact same structure as S3DIS. Then I ran the file "s3dis_from_scratch.sh" after preprocessing, but I got the following error. @JonasSchult

image image

windfill commented 10 months ago

I think the problem is caused by color. It cannot be a single color. In addition, have you reproduced the official dataset? I am more interested in the size of the final loss.

LinLin1031 commented 10 months ago

I think the problem is caused by color. It cannot be a single color. In addition, have you reproduced the official dataset? I am more interested in the size of the final loss.

Thank you for your comment!

I think you have hit the nail on the head. My custom dataset sets all point colors to be the same for each instance instead of the true RGB. I'm going to try again next by assigning random values to each point color as you suggested ( because I can't get the true RGB for each point).

As for your comment about final loss, I haven't gotten the results you're interested in because I haven't gotten into training well yet.

Lizhinwafu commented 8 months ago

I think the problem is caused by color. It cannot be a single color. In addition, have you reproduced the official dataset? I am more interested in the size of the final loss.

Thank you for your comment!

I think you have hit the nail on the head. My custom dataset sets all point colors to be the same for each instance instead of the true RGB. I'm going to try again next by assigning random values to each point color as you suggested ( because I can't get the true RGB for each point).

As for your comment about final loss, I haven't gotten the results you're interested in because I haven't gotten into training well yet.

Are you OK now?

tangze5258 commented 5 months ago

I think the problem is caused by color. It cannot be a single color. In addition, have you reproduced the official dataset? I am more interested in the size of the final loss.

Thank you for your comment!

I think you have hit the nail on the head. My custom dataset sets all point colors to be the same for each instance instead of the true RGB. I'm going to try again next by assigning random values to each point color as you suggested ( because I can't get the true RGB for each point).

As for your comment about final loss, I haven't gotten the results you're interested in because I haven't gotten into training well yet.

Have you solved the problem? how to set RGB if my data only intensity value ?