Danila Rukhovich
Danila Rukhovich
if you are not able to reproduce this, i can try to make minimal reproducible docker image
thank you again for your response. i also investigated that the problem is about the polygon with zero area. what i expect is to get an exception, but not like...
Thank you, i didn't know about this `clean=False`. Believe the memory leak will be fixed in future releases of cadquery or occ.
It is extremely strange as i just built the docker image with this dockerfile and everything is fine. By the way, why your directory is `/workspace/mmdetection3d` and not `/workspace/oneformer3d`?
You need point cloud data in [[x, y, z, r, g, b], ...] format and superpoint clusterinzation of it. Unfortunately this second part is not straightforward and can be done...
It is possible to initialize your custom dataset class even without `.pkl` files. They contain mainly just list of paths to point clouds and annotations. But we also provide scripts,...
Looks like `kitti_mm_dbinfos_train.pkl` is not the same with `kitti_dbinfos_train.pkl`. Are you going to provide this file or its generating script? Thanks.
Did you try the advice about `topk_insts` in our readme?
Actually, i don't remember, but `topk_insts=100` sounds good. I think, `inst_score_thr` may be increased easily up to 0.01, or even 0.05.
I think it is not an issue of oneformer3d, and the answer may be found in ScanNet [issues](https://github.com/ScanNet/ScanNet/issues).