Hello!

Thanks for open-sourcing this amazing work! Now I have a question about the figures in the paper.

Could you tell me how to get the visualization results in Figure-5 of the paper? Specifically, how to draw the transparent sphere with a discrete point cloud inside the sphere? Which tool did you use to render the point clouds to get that visualization results?

Thank you so much for your help!

what tools did you used for plot those figs?

Hi,

Thank you for the good paper, and for sharing the code.

I'd like to know if you have done a performance analysis in term of execution time. For example, how much does it take to train the classification or the segmentation method, or how mush does it take to do one iteration over the data?

Thank you in advance!

Hello @j1a0m0e4sNTU ,
I successfully executed this program, but I have some questions about the allocation of datasets.
Under normal circumstances, the training dataset and the validation dataset are carried out together.
And in your program it seems that the test dataset is used for training and also for testing.

# classification/main.py
# Line 55~72
if args.mode == "train":
    print('Trianing ...')
    train_data = ModelNet_pointcloud(args.dataset, 'train', transform= transform)
    train_loader = DataLoader(train_data, shuffle= True, batch_size= args.bs)
    test_data = ModelNet_pointcloud(args.dataset, 'test', transform= transform)
    test_loader = DataLoader(test_data, shuffle= False, batch_size= args.bs)

    # I have a problem here
    manager.train(train_loader, test_loader)

else:
    print('Testing ...')
    test_data = ModelNet_pointcloud(args.dataset, 'test', transform= transform)
    test_loader = DataLoader(test_data, shuffle= False, batch_size= args.bs)
    
    # I have a problem here 
    test_loss, test_acc = manager.test(test_loader)

    print('Test Loss: {:.5f}'.format(test_loss))
    print('Test Acc:  {:.5f}'.format(test_acc))

Hi,

Thank you for the great publication.

In fact, I'm a bit surprised that the results degrade if you rotate the objects, because rotation is a rigid transformation like the shift. If you rotate, the neighbor of the points doesn't change, as well as the distance between them, so how can you explain this degradation?

Thank you in advance for your answer!

dear，when add the -output in the train.sh or test.sh, I meet the problem. Could you help me? Thanks.

epoch 1 step 1000 | avg loss: 1.747 | miou(c): 0.440 | miou(i): 0.563
epoch 1 step 2000 | avg loss: 1.232 | miou(c): 0.498 | miou(i): 0.631
epoch 1 step 3000 | avg loss: 1.001 | miou(c): 0.532 | miou(i): 0.663
Traceback (most recent call last):
File "main.py", line 67, in
main()
File "main.py", line 56, in main
manager.train(train_loader, test_loader)
File "=/Graph/3dgcn-master/segmentation/manager.py", line 118, in train
test_loss, test_table_str = self.test(test_data, self.out_dir)
File "/Graph/3dgcn-master/segmentation/manager.py", line 151, in test
self.save_visualizations(out_dir, cat_name, obj_ids, points, labels, pred)
File "/Graph/3dgcn-master/segmentation/manager.py", line 81, in save_visualizations
visualize(point, label, gt_fig_name)
File "/Graph/3dgcn-master/segmentation/visualize.py", line 26, in visualize
ax.axis('off')
AttributeError: 'NoneType' object has no attribute 'axis'

Hi, thank you very much for your contribution to this project. I have a question to ask: what does your project hardware look like. What is the graphics card memory for the training project? thanks

Hi,

Thanks for providing the code. I was trying to setting up this repository with my another dataset where a single point cloud has 30k points. But in forward pass in first layer itself ( gcn3d.Conv_surface) the data value are getting nan at some positions , and resulting in loss as nan. So I wonder where I am going wrong ?

Looking forward to reply.

Thanks,
Charvi

Thanks for your wonderful paper!
Could you please provide the implementation of 3D-GCN with tensorflow? Thanks a lot!

I tried to recreate segmentation results but could achieve only
Best mIoU(c): 0.811, Best mIoU (i): 0.846

Please let me know if any changes need to be done to achieve the results in the paper
i.e. mIoU(c): 0.821, mIoU (i): 0.851

Thanks in Advance.

Congrats on the paper.

I downloaded the code and trained for classification as suggested in the Readme file.
However, upon the test, I could only get a 90.8% classification accuracy.

Please let us know how to achieve 92.1% accuracy on classification as stated in the paper.

Thanks in advance.