jiashunwang/neural-pose-transfer

Neural Pose Transfer by Spatially Adaptive Instance Normalization. In CVPR 2020

neural-pose-transfer's Issues

How to test the model if the vertices of identity mesh and pose mesh are different?

Hi @jiashunwang
I want to try your model on human mesh with different vertex numbers. I change the num_points in model_maxpool, but dimension error will be reported when performing cat operation. Then I try to reduce the point of the one with more points to be the same as the other in the random_sample. Then the code can run. However, the results are not very ideal.
Could you please give me some advice? Thank you!

random_sample = np.random.choice(6890,size=6890,replace=False)
random_sample2 = np.random.choice(27554,size=6890,replace=False)

Facial expressions and fingers

Hello @jiashunwang & team,
Great paper. Thanks for sharing the code.

Wish to know if the model will be able to do these:

1. transfer facial expressions
2. transfer exact finger positions (all 10 fingers)

Thanks,

test results are bad when use other datasets

Hi @jiashunwang ,thanks for sharing this wondeful work, when I test the pretained model on the datasets you provided, the result looks normal, but when I test the obj file from other smpl based model like VIBE, the results are bad.
I wonder whether it's because of the model's generalized ability or there exist some details that I missed?

Question about Permutation Invariant Pose Feature Extractor

Thanks for your great work!
But I wonder why the model works on target and source mesh with a different vertex order. The feature F_pose extracted from the Encoder seems to be different on the same pose mesh with different vertex order. Also, F_pose is not spatially aligned with the vertex of the identity mesh. Does Pose1(order 1) + Identity1 and Pose1(order 2) + Identity1 outputs nearly the same result? Any insights on why it works on different vertex order inputs?

Could you please upload the model for horses?

Dear @TianyunZ @walsvid @jiashunwang
I just run the demo.py with two horses. The model for a horse is different from the model link you provided.
Could you please upload the model for horses?

Thanks!

The results of supervised_list_obj.txt and unsupervised_list_obj.txt are very close.

I just tested the models you provided and found that the results of supervised_list_obj.txt and unsupervised_list_obj.txt on both maxpool.model and original.model are very close. Moreover, they are much better than the results in Table 1 of your paper. So I think the two txt files may not correct, can you help me figure this out? Many thanks.

Can not find the Point-wise Mesh Euclidean Distance measurement code

Hi, thanks for your good work, but we cannot find the protocols to measure the performance, especially the code of Point-wise Mesh Euclidean Distance for supervised and unsupervised (only list)

Could you provide one?

Can't access datasets

The website is saying that "Under Maintenance", so we can't access datasets. Could you upload the datasets to Baidu Yun or Google Drive?

Permision

You don't have permission to access /fuyanwei/download/NeuralPoseTransfer/data/filelists/supervised_list_obj.txt on this server.

Will you share the pre-trianed model and test cases?

Would you please sahre the pretrained model and the demo code? I hope to test this model on my own data. Also, how to set the training data? Looking forward to your reply!

the topology structure of ground true keep the same as the identity mesh

Hi, I recently read your paper related about neural pose transfer. After read the codes, I am confused about why the generated mesh should keep the same topology structure as the identity mesh provided, why not pose mesh?

How to visualize the data? Can you provide more details?

no permission to the pre-trained models

I've tried to download the pre-trained models in http://www.sdspeople.fudan.edu.cn/fuyanwei/download/NeuralPoseTransfer/ckpt/,
but I met the forbidden: "You don't have permission to access /fuyanwei/download/NeuralPoseTransfer/ckpt/original.model on this server"
Have I missed any authentication steps?

Visualization of the different index order of meshes

Hi Jiashun,

Could you please share how you visual the meshes with different vertex orders? Thank you.

Reg. SMPL based dataset

Hi @jiashunwang,
I understand that we can use the .obj files created by SMPL based models as dataset.
Do let me know if there is any constraints like entire 3D object should be visible, etc.

Ca you share sample dataset that you used for training? That will help me understand how to create dataset for training.

Thanks