czq142857 / ndc Goto Github PK

cloud points:

undc result:

cloud points:

undc result:

get mesh but there is no points,
anything i can try to solve such problem ??

thanks

Hi, great work and thanks for sharing!
I meet some problems when I tested this method. I found that when reconstructing a surface with UNDC using a point cloud as input, the output always surffer from wrong face normals and obvious "zigzag" issues, even I use the default input parameters(python main.py --test_input examples/tshirt.ply --input_type pointcloud --method undc --postprocessing --point_num 8192 --grid_size 128
).
I don't know if this is common? In addtion, are there any suggestions to improve the reconstrution quality? Thanks a lot!

Hello, ZHIQIN CHEN
Thank your for your kindly share!
I noticed that the out put of NDC and NMC is total different. For I see the output of NMC is [M,N,K,5] for bool and [M,N,K,51] for float just as explanation in your paper. But in NDC, the out put of float part is [M,N,K,3]. I am not very sure for your encoding for dual countering. May I ask if you have another document for detailed explanation of the NDC project.
It will my appreciation for your responce.

Great work! I was inspired by your work. Now I am trying to do some extensions based on your work.
I hope to know the original code of VOXGen, SDFGen, and IntersectionXYZpn? Where do you find them?

Thanks,
Ruowei Wang

Hi! Thanks for sharing this great work.

I installed the repo and tried to reconstruct the Mobius example you have, using your weights/data/commands from the also provided low-res point cloud (1024 pts). The output however has many holes and seems off compared to the one in the teaser.

Is this expected? If so, can you please point to the steps needed to reconstruct the pc with similar quality to the one shown in your teaser?

Hi. The paper mentions that a lot of stuff can be found in the "supplementary material". But I cannot find this easily anywhere. Can you point me to this material? Thanks!

Table 1 in your papers says that training time < 12hours per net work.But it takes about 19 hours when I train the undc float network with point cloud input(gpu rtx A6000). And the batch size must be 1 in data loader,I want to know why.Because actually the data loader and network can be changed to process multiple shapes in one batch.

Thanks for your great work!
But when I test undc on pointcloud input, it seems to be too sharp. I use meshlab to visualize the output mesh model.

Hello @czq142857

Thanks for sharing the project. Does the current implementation support taking point cloud as inputs (testing UNDC)? I know there are related examples shown in the paper. But I don't see script/instruction for point cloud except for SDF and voxel.

Hello,

I observe that the GT SDF provided has some faulty reconstruction. In the below picture, I visualise and compare interior of the same shape. Clockwise from top-left are, GT, Marching Cubes on provided SDF, Dual Contouring. As you can see, the GT does not contain the 4 pillars which are present in the other three shapes. What is the reason behind this?

Edit: This is shape 00007937.obj from the test set.

Hello, Dr. Chen，
Thank you very much for your outstanding contribution to point cloud surface reconstruction.
In order to train the undc_with_noise model better, I see that there are about 200,000 samples after data augmentation.
I tried to train the model using the source code, but the training process was slow, probably due to the large amount of data.
I would like to ask you, what hardware did you use for training? How long did it take you to train the model?
Or do you know of a way to speed up model training?

Hi! @czq142857. Thanks for sharing the great work.
I wonder know if the training data for noisy point cloud UNDC is the same as others input(using ABC), but adding additional data augmentation. Matterport3D's point cloud is just for testing not training.
And what's the key factor that you think makes the model generalize well on real data?

Hello,

I'm trying to overfit the network to a single shape and I find that the training results are significantly different from the pre-trained model. Attached below is the discrepancy I'm referring to. What could be the reason?

Edit: Left in the figure is the pretrained model of UNDC Point cloud (both bool and float). Right uses the ground truth vertex positions and the connectivity is what I overfit to the single shape and try to predict the result on the same. (done for a sanity check)