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

How is the performance on S3DIS dataset evaluated? #124

Open jerry3chen opened 1 year ago

jerry3chen commented 1 year ago

Hi. According to your paper, you used 6-fold cross validation. To my understanding, it means that you run it 6 times, each time keeping 1 area as test set and the other 5 areas as training set.

If I am correct, here are my questions:

  1. what data was used as validation set? How did you determine when your model started overfitting.
  2. In your paper, are the other methods also evaluated in the same way? I thought whatever their training strategy is, they always use area 5 as the test set.
  3. bonus question: I am trying to run your code on 3090ti but I will be out of memory whenever I do evaluation on full scene, what can I do to make it work on my GPU?

Thanks, Jerry