ScanObjectNN

[dtld123]

Thank you for proposing such an inspiring method. However, my performance on the ScanObjectNN dataset is quite poor. I suspect it might be related to the number of points (2048 points). Do you have any suggestions for improvement?

[eliphatfs]

Are you matching the reported performance in paper?

[dtld123]

Yes, I can only get about 10% accuracy on the ScanObjectNN dataset. I don't know where the error is.

[eliphatfs]

Are you using the provided ScanObjectNNTest class for data loading?

[dtld123]

Thanks for your reminder. I tested the results using the cat_name_pt_feat.npy and xyz_label.npy files you provided. The OpenShape-SparseConv + ScanObjectNN result is 48.93, which is slightly lower than the 56.7 reported in the paper. I'm not sure if this is normal. I'd also like to know why the results are so poor when I use the official ScanObjectNN test set directly. Did you perform any changes to the test set?

[eliphatfs]

There are multiple splits for the scanobjectnn dataset. I forgot which we used now but I can check shortly.

[dtld123]

I understand that the simplest slice is main_split_nobg/objectdataset. Could you please help check it?

[dtld123]

This is very strange. The number of test sets given by OpenShape is 2890, but none of the partitions of scanobjectnn have this number.

[eliphatfs]

It should be OBJ_ONLY with 581 shapes.