papagina / meshconvolution Goto Github PK

Hello. I am a University student studying how to use 3D Human data in Korea. Thank you so much for posting such a good source code. Even though you uploaded it well, I am not able to perform it properly right now. Sorry.
I tried to perform GraphSampling using the current DFAUST data, but an error appears during cmake.
I really don't know how to fix the error. I searched the internet and couldn't find it.
cmake doesn't work well and I'm not getting information about "ConnectionMatrices" right now.
So maybe I can get connectionMatrices like MeshConvolution/train/0223_GraphAE27_compare/?
I'm really sorry. please.

Both this repository and VCMeshConv (by "facebookresearch") claim to provide the code for this paper

However, this repository employs an MIT license, and the VCMeshConv repository employs a non-commercial Creative Commons license

Could the author provide some clarity on the conflicting licenses?

Hi,
I got an error:
HTTP request sent, awaiting response... 403 Forbidden
2021-12-25 14:11:39 ERROR 403: Forbidden.

validation data: https://meshfullyconvolutionalautoencoderdata.s3.amazonaws.com/sakuratrunk/validation.npy

Would you help me to download it?
Thanks in advance
Happy new year!

Hello,

I am curious to know given a 2D image of a human, is your model able to do 3D lifting of the human mesh similar to what SMPL-X and ExPose do and if yours produce more expressive results?

Also, which metric do you use to compare your results with SMPL-X/ExPose 3D mesh body models?

I get the following segmentation fault when I follow the instructions on the default template.obj:

$ ./GraphSampling
Load mesh ../../data/DFAUST/template.obj
#############################################################
## Create pool and unpool layers ############################
## Add Pooling 0
Start setting connection map from mesh
Start setting must_include_center_lst from mesh (the red points)
There are 6 red points.
stride: 1, radius pool: 1, radius unpool: 1Center points number: 6890
## Add Pooling 1
stride: 2, radius pool: 2, radius unpool: 2Center points number: 1925
## Add Pooling 2
stride: 1, radius pool: 1, radius unpool: 1Center points number: 1925
## Add Pooling 3
stride: 2, radius pool: 2, radius unpool: 2Center points number: 400
## Add Pooling 4
stride: 1, radius pool: 1, radius unpool: 1Center points number: 400
## Add Pooling 5
stride: 2, radius pool: 2, radius unpool: 2Center points number: 54
## Add Pooling 6
stride: 1, radius pool: 1, radius unpool: 1Center points number: 54
## Add Pooling 7
stride: 2, radius pool: 2, radius unpool: 2Center points number: 7
#############################################################
## Save pool and unpool connection matrices in npy ##########
save ../../train/0422_graphAE_dfaust/ConnectionMatrices/_pool0.npy
output point num: 6890; max_neighbor_num: 21
Segmentation fault (core dumped)

The issue seems to stem from &pool_neighborID_lst_lst [0] on line 114 in GraphSampling/meshCNN.h:

cnpy::npy_save(save_path+"_pool"+to_string(i)+".npy", &pool_neighborID_lst_lst [0], shape_info, "w");//"a" appends to the file we created above

Has anyone else encountered this problem and does anyone know the fix?

Hi there, would it be possible to get license information for this repository please? Thanks!

Hi, thank you so much for an excellent paper and code! I'm looking at using this network as part of a volume to mesh generative model, but the mesh data I have currently is non-registered. Is there potential to apply this network to non-registered meshes using the convolution/pooling operations you have developed? What components make mesh registration necessary if not? Thank you!

Hi,
thank you for your paper. I have one question regarding training on a custom dataset. I have several stl geometries of varying mesh resolution. I use one of them as template to get the connection matrix. Can i just sample n-points from each geometry to train the network or should the point clouds have a similar correspondence? Since you say in the paper that network is template-free, does the point order and point location different between the geometries make any difference in training?
Thanks a lot.
Regards

Hi,
I noticed that your code allows to autoencode for multiple components, which I interpreted it as it allows to make some kind of "sharing subspace" between components, by using the same model parameters.
Is it possible to apply the trained model, which is trained with multiple components, on separate components?