search

InterDigitalInc / TearingNet

Source code for the paper "TearingNet: Point Cloud Autoencoder to Learn Topology-Friendly Representations"
Other
26 stars 5 forks source link

visualization software #4

Open zhaziyaoyangguang opened 1 year ago

zhaziyaoyangguang commented 1 year ago

Hello, I think this article is very interesting and meaningful. May I ask what visualization software is used for the data in your paper? I hope you can tell me, thank you.

zhaziyaoyangguang commented 1 year ago

Hello, I would like to ask you another question. The dataset I used was cad. After the training of foldingnet and tearingnet in two stages according to the script you gave, I used the reconstruction script for testing, and found that the result of CD was similar to the original text. However, the results of opening the reconstruction with cloudcompare are not good enough to show the effect of reconstructing the point cloud. Besides, the generated 2d grid does not have obvious tearing effect. Could you please help me to solve it? image

image

Looking forward to your reply. Thank you!

jpangidcc commented 1 year ago

Hi @zhaziyaoyangguang thanks for your interest and your comments about our work! We have not published our visualization code, but for your information, the visualization is performed with Open3D and we render each point as a sphere for better visualization.

As for the training of the TearingNet, we first pretrain the encoder (PointNet) and the Folding decoder together, with a loss function that downscales the second term of Eq. (4) in the paper. Next, we train the entireTearingNet using a smaller learning rate via the complete CD loss (Eq. (4)). See the Section III-A of the supplementary material for more details.

atomishcv commented 1 year ago

hi @jpangidcc it is a great job. I am currently able to reconstruct multiple objects, but the 2D torn grid I output does not look like the one given in your paper. Could you please tell me the specific steps for making a 2d torn grid? image

jpangidcc commented 1 year ago

Hi @atomishcv, in order to reproduce the torn grid pattern, and more importantly, to have better reconstruction quality, it is necessary to use a two-stage training strategy. Particularly, we first pretrain the encoder (PointNet) and the Folding decoder together, with a loss function that downscales the second term of Eq. (4) in the paper. Next, we train the entireTearingNet using a smaller learning rate via the complete CD loss (Eq. (4)). See the Section III-A of the supplementary material for more details.