PointCNN The performance of PointCNN on TU-Berlin

I ran the code of PointCNN on TU-Berlin (using a single Tesla V100 GPU), strictly following the code and hyper-parameter settings based on the released github version. The validation accuracy of PointCNN reached 60.% at 40K iteration, but was no longer increased, while the reported result is 70.57% in the NIPS version. I visualized the training/validation loss and training/validation accuracy as follows: screenshot from 2019-02-19 09-35-24 I found that the the loss did not converged to a sufficiently small value. I also tried several small initial learning rates such as 0.001, 0.0001, but got similar results. Do I have missed something that is important to the performance of PointCNN? @burui11087

Feb 19 '19 12:02 jiaxin19cvml

Hi @cv2drpepper

The network performance you trained on TU-Berlin looks very strange, I need some days to rerun expriment on TU-Berlin to verify the issue you reported. BTW, could you post your computer environments such as CUDA,python,TF version.

Thanks.

Mar 28 '19 08:03 burui11087

Hi @burui11087

Thanks for your response. My computer environments are: CUDA 9.0; CUDNN 7.3.1; python 3.6 and tensorflow-gpu 1.9.0.

Thanks

Mar 28 '19 09:03 jiaxin19cvml

Hi @jiaxin19cvml

I find that we augment testing datasets when preparing train/val/test datasets so that acc. is just 60%. I will update code soon in these few days.

Thanks

Apr 04 '19 05:04 burui11087

That's fine. Looking forward to your update code. @burui11087 Thanks~

Apr 04 '19 06:04 jiaxin19cvml

Hi, I try to process the test dataset without augmented. Now it has 6666 samples of test. Then I try to reproduce the result you report, still can't get 70.57, but get ~67. It seems not increase for a long time, is there anything wrong? Here is the screenshot of tensorboard:

May 29 '19 23:05 Leiay