azuki-miho

You'd better ask the administrator of the Semantic3D benchmark or create a new classifier

I just have a try and it doesn't succeed. Maybe there are some problems in the Semantic3D server, you'd better ask the administrator of Semantic3D.

Thanks for your question. Even though we only test our code in the Semantic3D reduced-8 dataset, I can offer some hints to solve this problem. This problem seems to occur...

It takes about 2 days in my memory.

I think the training time of KPConv is also about 2 days for the Semantic3D dataset. The RFCC generation and extra supervision don't introduce much extra training time.

Training over multiple GPU is not supported currently. I suggest using 1080Ti or 2080Ti for training. If you only have 2 GTX 2080, I recommend you lower down the batch...

It seems you utilize the KPConv backbone, whether it is possible to run the code with naive KPConv?

As RFCR is relatively simple to implement and only a few lines of code should be inserted into the backbone, whether it is possible to first run the code and...

Actually, I think you should first check where the problem from, the Dataset, the KPConv backbone or the RFCR module, and the solve the problem.

As the KPConv original network, whether it is possible to add the RFCC generation, RFCC supervision, and Feature Densification separately and locate the problem.