sw-gong/spiralnet_plus

SpiralNet++: A Fast and Highly Efficient Mesh Convolution Operator (ICCV-W 2019)

spiralnet_plus's Issues

where can we find the code for shape reconstruction?

where can we find the code for shape reconstruction?

Thanks!

Working with meshes having different number of vertices?

Thanks for sharing your code.

I have a dataset of 3D meshes having a different number of vertices.

The size of convolution layers depends on indices parameter and they are different for each mesh so Is there a way to train SpiralNet on such dataset?

Regarding how to call the model on custom dataset

Hi,
I want to run your model on my custom dataset. I'm struggling with how to run the model with the mesh template which I have. Can you please help. Thanks in advance

How to incorporate auxiliary input information

Thanks for awesome work.

I'm trying to use SpiralNet++ for CFD analysis (something like this paper but for 3D) and getting good results.

I would like to know what's the best way to incorporate auxiliary information (other than input coordinates and spiral sequences) in the network.

For example, I would like to introduce input freestream velocity (x,y, z-direction) and angle of attack (scaler value) to the network.

Many Thanks.

Shape Correspondece

Hi, thank you for your work.
I have tried your code about shape correspondence on FAUST and the results are close with that in your paper. However, I noticed that you did not shuffle the vertex orders, so I want to know why we need a GCN to learn correspondence since the vertices of meshes in FAUST are already in one-to-one correspondence.
And I have tried to shuffle the vertex orders then Spiralnet++ did not work. So I am confused and want to ask for your advice. Thanks a lot.

Working Colab Notebook with updated dependencies!

Hello,

This is more of an observation than an issue. I spent a considerable time setting up the environment and getting everything to work correctly (especially the CUDA versions of Pytorch and Pytorch-Scatter and Pytorch-Sparse). I have everything working in a Google Colab Notebook. But I would like to make a Docker Image file with all the specifications for easier onboarding. (And to get it working on my university servers) Although the 13 GB RAM is insufficient for running python -m reconstruction.main to completion, at least we know it's working.

Also, if you are convinced about the dependencies, I could update the list of dependencies in the README. Let me know if you want me to put in a pull request :)

I have installed the environment on Google Colab successfully with the following specifications -

• PyTorch 1.9 with CUDA 11.1
• PyTorch Geometric 1.9 with CUDA 11.1 (Followed this handy script for the correct scatter and sparse CUDA versions)
• OpenMesh 1.1.6
• MPI-IS Mesh installed from source (instructions in the notebook)

I'll update here once I have a working Dockerfile with the above environment. Cheers :)

When will the code be released?

I'm using spiral convolution for one of my project and it works well. So I would like to plug in spiral++ and see how it goes. Will the code be released this weekends?

How is the information propagated in the process of pooling?

As I understood, in the encoders' forward of the reconstruction model, pooling operation only preserves features of the predefined (by the transform.pkl file) vertices. Since there is no averaging or maxing out, how is the information from the neighboring vertices propagated into downsampled representations of the mesh? By participating in the convolutions before poolings only?

Thank you!

constant serialization

Hi,

Thanks for your code and paper. It looks concise and great. I have a quick question here and hope you could give me some directions.

In the FAUST experiment, the spirals are defined by the first data and keep unchanged during the training. In geometry, this serialization doesn't stand for spirals for other data. Is there any explanation for it?

Thanks!

Pretrained model for reconstruction network?

Hello Shunwang Gong,

I am trying to train the reconstruction model on a Tesla K40 (12GB memory) but unfortunately running out of GPU memory mid-way training. I'm training the network with the CoMA dataset as given in the code (more than 20k faces).

Could you help me with some strategies to combat this issue? I do have a cluster of GPUs available at my disposal but since the paper doesn't talk about any resources required to train the networks, I thought I would ask you first.

Or perhaps if it's possible, could you please share the pretrained model for reconstruction? I would like to use the encoder network on some of my in-house dataset.

Thank you,
Niraj Pandkar

FAUST autoencoder generates really noisy outputs

My network architecture :

• latent_channels : 256
• out_channels = [16, 32, 64, 128]

Is there anything else that I need to set for FAUST Autoencoder training?