about the fine tune? about pointcnn HOT 5 CLOSED

Hi @lfxx,

I think there can be two potential directions:

1. Synthesize data: if your testing data is not too far from what you can synthesize, then I think the model trained on synthetic dataset should work reasonably well on real data.
2. Scan and label real data: if 1 doesn't work, probably you have to go through this approach. Similar to how ScanNet was created.

from pointcnn.

采用第二种方法的话，还是需要在scannet数据集上做预训练吧，然后再在自己的训练集上调优？

from pointcnn.

Maybe, I cannot tell...

It depends on the object types you want to work with, as well as how much data you can scan/label.

from pointcnn.

1.The datasets made on my own is small,so is it a good idea to get the net pretrained on scannet datasets?
2.我自制的数据集肯定是比较小的那种，所以感觉还是需要预训练一下，这样效果可能会比较好?

from pointcnn.

Try it ;-)

from pointcnn.