Thanks about the code, I need some information about the output files. For example for the chair class of pcn there are 5 folders as pcd0 to pcd3 and pcdc with many files in each of them. First what are they and why are they in 5 different files and second, what kind of tools do you suggest for visualizing and considering them? I found that all of the output .h5 files have the same data values!

Hi, I trained the model based on the PCN dataset. The average L1 chamfer distance is 5.1510^(-3). The pre-trained model you provided is 5.0710^(-3), which is better than the data 7.21*10^(-3) in the paper. Except for the dataset path, I did not modify your code. Is this result correct?

Thank you very much for releasing the code. I'm seeking a good visualization tool for point cloud. According to your paper, the visualized point cloud is beautiful. Could you please release the code for visualization?

Hi,

Thanks for your work!

I noticed that your public shapenet training data is saved in 'train.lmdb'. May I ask how you read shapenet data from lmdb file？

In utils/data_loaders.py, the dataset class always return the partial point cloud and grount truth point cloud. But when I try to inference using pretrained models, an error occurs. This is because for testing split, there is no ground truth data, i.e. /path/to/datasets/Completion3D/test/gt doesn't exist. The dataset is download from the link you provided. So, could you please provide the corresponding ground truth of testing split of completion3d?

Hi!

I just wanted to report what I think might be a mistake in using the PCN test data not only for testing, but also for validation. Shouldn't utils.data_loaders.DatasetSubset.TEST in the following rather be utils.data_loaders.DatasetSubset.VAL..? https://github.com/AllenXiangX/SnowflakeNet/blob/93e7151610765e7e2b41ace2d03c8750f0b6c80c/core/train_pcn.py#L36-L37

Regards, Lucas

Hello, authors! I would like to know how you obtained the L2 Chamfer distance in Table 2 of your paper. After executing your code on the c3d dataset, I obtained the following outcome, which seems to be the L1 Chamfer distance.

============================ TEST RESULTS ============================
Taxonomy        #Sample ChamferDistance
02691156        100     0.2252
02933112        100     1.2307
02958343        100     0.6126
03001627        100     1.0656
03636649        100     1.1419
04256520        100     0.8829
04379243        100     1.2952
04530566        100     0.5210
Overall                 0.8719

Epoch  -1       2.8626  1.6381  1.1422  0.8719  0.0540

Hi, I want to use myself training dataset, dose it possible? Could I set each of these point cloud dataset to a fixed size of 200000 points for-example, instead of 16348 points which you had set for pcn dataset? The best Ema

Hi, I noticed that you are scaling the point clouds by 0.85 during validation - #. Wouldn't this reduce the chamfer value as the whole object is now compressed to a smaller size? To confirm this I computed the chamfer distance with and without it on the validation set using the provided pre-trained model (c3d).

Validation transforms as is - CD: 0.000872 (is as expected and close to the reported numbers on the test set) Using the test set transforms during validation (i.e only ToTensor()) - CD: 0.001968 (more than twice the previous case)

It would be really helpful if you could clarify the same. Thanks

hi, I wonder how do you define your training strategy (e.g. batchsize in pcn case and hardware infos) and how much epochs does it take to converge? Looking forward to your reply!