A question about MS-SVconv code.

[QWTforGithub]

Thank you very much for releasing the code on torch - points3D. Now I want to further work on your work, I hope to have clean code on MS-SVconv, can you publish it? It is very troublesome to modify the part of MS-SVconv directly in torch-points3D.

Looking forward to your reply.

[humanpose1]

Hi,
Thanks for your interest in MS-SVConv. The code of the model is here if you want to play with it. But indeed, you'll need the framework to instantiate the model (this function). You need the dataset in order to automatically instantiate the model with the right input size.you need the config file .(for example this). I hope it can help you.
I think that you don't have to clone the repo to use these functions. You can install tp3d with pip.

Actually, you're right. When I'll have time, I will publish a simple version of the code of the model here.

Don't hesitate if you have other questions.

[QWTforGithub]

Thank you for your replying.I look forward to your release of a simpler version.

Best wish.

[QWTforGithub]

Hi,
Thanks for your interest in MS-SVConv. The code of the model is here if you want to play with it. But indeed, you'll need the framework to instantiate the model (this function). You need the dataset in order to automatically instantiate the model with the right input size.you need the config file .(for example this). I hope it can help you.
I think that you don't have to clone the repo to use these functions. You can install tp3d with pip.

Actually, you're right. When I'll have time, I will publish a simple version of the code of the model here.

Don't hesitate if you have other questions.

By the way, I want to use MinkowskiEngin instead of Tensor_sparse to build a network for training. Is the result of MinkowskiEngin worse than Tensor_sparse? Or is there just a difference in speed that doesn't affect the recall results.

I looking forward your reply, thank you.

[humanpose1]

Hi @QWTforGithub ,

You can use ME instead of Torchsparse, it will not affect the result. It is the same convolution.There can be a difference in speed depending on your version of ME or Torchsparse.

Best,