weiguangzhao
Request for Custom Dataset Training Guidelines
Hi, thank you for the work. I am interested in training custom datasets using your method. Would there be provide sooner? Thank you.
Thanks for your attention. We will provide custom data training guidance as soon as possible. In general, we may need approximately one month for debugging.
Thank you very much for your prompt response. At the meantime, would there be any recommend way for me to run training before releasing of the guideline? like the dataset would be in scannet format or s3dis format.
There are three points you may need to pay attention to:
1.Convert all data to the npy format, including XYZ [N, 3]; RGB [N, 3]; Semantic Label [N]; Instance Label [N]; and set the labels for invalid or noisy points to -100. For this part, you need to refine the dataset.py file.
2.Finetuning the hyperparameter raduis and min_pts in the config.py&config_test.py. We recommend standardizing the coordinate data to the meter level, then setting the radius to 0.04, and min_pts to 1.5 times the original point density.
- Finally, you need to modify the number of semantic categories
Feel free to ask if you have any questions.