guochengqian/PU-GCN

Pretrained-Model

stevenlin510 opened this issue · 6 comments

stevenlin510 commented

Hi, thanks for amazing work, could you provide the pre-trained model you have in your paper ? Thanks !
And another question is that did you use monte-carlo sampling to generate the test samples (not ground turth) or use poission-disk sampling ?

guochengqian commented

simply train the code on the provided dataset for less than 3 hours, you will have almost the same result as the paper.

guochengqian commented

okay. i will release it next week. i am busy with other stuff these days.

  • 👍2
stevenlin510 commented

Thanks !! I just wondering how to get the excellent quantitative results on PUGAN's dataset, it seems that the test samples which PUGAN using were sampled by monte-carlo sampling(2048), and what you might using poisson-disk sampling to generate test sample(2048) right ?

guochengqian commented

Thanks !! I just wondering how to get the excellent quantitative results on PUGAN's dataset, it seems that the test samples which PUGAN using were sampled by monte-carlo sampling(2048), and what you might using poisson-disk sampling to generate test sample(2048) right ?

yes. for comparasion on PU-GAN's dataset, it was not that fair, since the input configuration was not the same.

guochengqian commented

FYI, the pre-trained models are released.
One can also train the models by themselves on PU1K dataset. We support training PU-Net, MPU, PU-GCN. The results are reproducible. We trained 10 times and got close results.

lai-pf commented

thanks for sharing the code, which is really generous .
And may I ask the training settings about PU-Net,MPU,PUGCN model, should I training all these model use the same setting like the default setting you set? or each model has different settings in training?
thanks again