The provided pretrained model (https://drive.google.com/file/d/1fj84kyD9CGT8j61IW-xSWZ5q4q5IpoYx/view?usp=sharing) seems to be broken, could you please upload a complete one?

Thanks for the code, and the great work. Hope you can explain some training details.

1. Whether data augmentation was used during training, the training setting in readme.md is different with supplementary Material.
2. Are there any setting that have a huge impact on the results? I rebuild the network and trained it for 100 epoches with initial lr 1e-3 and decay step 40, but both $L_{emb}$ and $L_{nnl}$ are worse than the provided pretrained model on the test set. Thanks for answer

Hi, thanks for the code, and the great work. But there are a lot of places in the paper that confuse me， hope you can clarify： 1：What means R∈O(K) in Equation1？ 2：Ac = Ag_c + E ? 3: What does ei mean in section 4.3? 4: Does Fl mean that each point has L=1+ dc + ds (dc=K)features in the sub_section Weighting sub-module？ 5: Equation 6 and Equation 7 are confusing, k<k' is different from reference【1，34】? 6: About Evaluation metrics seg_iou: the paper doesn't seem to mention how to predict membership matrix W.

Hi, Siming, thanks for the code and the dataset. I noticed that there are 4000 test samples in your test dataset, but I see that the test set written in your code is to take every 100th test sample as a test sample. I would like to ask whether the results in your paper are based on the full test set or the results of randomly selected partial samples for testing

Hi Siming， I wonder where is the primitive parameters information come in your dataset? I didn't find it in the origin ABC dataset, did you pre-process it from ABC dataset? And I also curious about the accuracy of directly predict primitive parameters from network, is it similar to the gt primitive parameters? Thanks in advance for your reply and wonderful code. It helps me a lot.~

Hello,

Thanks for making your code available.

I downloaded your pre-trained model which is abc_normal.tar from your given link, but I can not unzip that file. Can you check if the file is broken?

best. Mulin

Thanks for your great work!

I'm running into a problem in the training part, it reports CUDA out of memory but there;s no any other program running on the gpu I select, and the memory is obviously sufficient.

Here are the error message and the infomation of my gpu. I wonder how to solve this and looking forward to your reply!

Hi, Siming. I tried to eval the network on a server with 4 1080P GPU and 16 cpus with command: CUDA_VISIBLE_DEVICES=1 python train.py --eval --checkpoint_path=abc_normal.tar --val_skip=100 --input_normal=1

But I found that the mean shift clustering step is rather slow, it takes over 3 minites to compute the clustering of a single model. And the utilization of 16 cpu reaches to 100%. Do you have any suggestions to speed up this step? Thanks a lot.

Hi, Siming. Thanks for providing the dataset, can you tell me where I can find the correpondence between your data id and the id of ABC dataset? That would be very helpful. Thanks.

Hello,

Can you explain a little bit what is the meaning of each dimension of "T_param"?

I know it's the parameters of plane, cone, sphere and cylinder. I think you follow the parameters defined in SPFN, can you describe it more precisely? For example, the first 3 dimensions is 'a' of a plane or the first dimension is 'd' of a plane?

Thanks in advance.

Best. Mulin

Hello Siming,

I run 'python train.py --eval --checkpoint_path=./path/to/pretrained/model --val_skip=100' to test your code with your provided data and pre-trained model, the results are: 'feat_loss: 0.770812 miou: 0.384534 nnl_loss: 2.341513 param_loss: 0.048104 type_miou: 0.170333' which are far away from the results shown in the paper. And I also visualized several results which are not well segmented. I want to figure out where is the problem since I didn't touch your code and use the default parameters. Can you check the pre-trained model?

Thanks in advance. Best regard and have a nice day. Best. Mulin