rubenvillegas / cvpr2018nkn Goto Github PK

Thanks for your great work! I am currently studying this project but encountered a download problem. I also looked at other issues that others might have the same issue with me. However. it was about a year ago on 25 April 2019. Could you please check the link again? Thanks a lot!

Do you know any better way to download the FBX file from the website? should I download it one by one?

Hi, the autoencoder loss and cycle loss described in the paper should has a global optimal solution that is just copying quaternions and velocity, but why the GRU model learn something different?

The umich.box link is expired, and while #12 mentions a Google Drive link, it is expired as well...
Is there a way to obtain the test files ?

Thanks

I am trying to download the test data using the command ./datasets/download_test.sh, but the link https://umich.box.com/shared/static/5q39kgt2ftkfy352rr2lmwuzrhqjxwvy.gz cannot access. could you please put the test data on the Google drive or Baidu Netdisk? Thank you very much~

thanks for the work. i have an issue when downloading the blender_files in video generating. the shared files have been removed

thanks

Hi!

Thanks for sharing such an impressive job.

I have a question, if the network output is the velocity of the root joint, then where does the position of the first frame come from?

It seems that the initial position of the root joint comes from the ground truth as shown in test_online_retargeting_mixamo.py

outputB_bvh[:, :3] = gtanim.positions[:1, 0, :].copy()
wjs, rots = put_in_world_bvh(outputB_bvh.copy(), start_rots)
tjoints[:, 0, :] = wjs[0, :, 0].copy()

Looking forward to your kind reply.

Dear authors,
To run the codes, it seems it is required to have all the training samples. However, downloading and processing all the 1656 samples one-by-one can be really tedious. Would you please help us if you have any suggestion to facilitate the collecting of training samples. Sending a pre-trained model is highly appreciated

Sorry for the issue.
We are studying this project, and we are trying to use data other than Mixamo.
During testing, we found that there are three npy files for each bvh in the test data downloaded from download_test.sh, but we cannot process the model_ang.npy for any extra data by using prepreocess.py.
May you provide a preprocessing program used in preprocess test data?
It would be really helpful for us. Thank you.

Hi, it seems that mixamo adds some motion sequences? It can not been matched with your shared test set.
Would you like to share the train list, this can really help me download corresponding motion sequences. Thanks a lot!

I try to inference test data after training NKN with Adversarial Cycle Consistency.
Training data was downloaded from mixamo site, and converted to npy file using datasets/fbx2bvh.py and datasets/preprocess.py.

run command

CUDA_VISIBLE_DEVICES=0 python src/test_online_retargeting_mixamo.py --gpu=0 --prefix=Online_Retargeting_Mixamo_Cycle_Adv_beta\=0.001_gru_units\=512_optim\=adam_d_arch\=2_learning_rate\=0.0001_omega\=0.01_norm_type\=batch_norm_d_rand\=True_num_layer\=2_alpha\=100.0_euler_ord\=yzx_margin\=0.3_keep_prob\=0.9_gamma\=10.0

But I got this error.

output error code

[*] Reading checkpoints...
     Loaded model: EncoderDecoderGRU.model-50094
[*] Load SUCCESS
Testing: 0/185
Traceback (most recent call last):
  File "src/test_online_retargeting_mixamo.py", line 240, in <module>
    main(**vars(args))
  File "src/test_online_retargeting_mixamo.py", line 139, in main
    -4] = localA_batch[:, :step, :-4] * local_std + local_mean
ValueError: operands could not be broadcast together with shapes (1,120,62) (1,1,66)

I downloaded test data using datasets/download_test.sh.
These shapes of the test data loaded npy file is (120, 66).
I think this code intended shape of the test data is (120, 70), but this data seems not to include root_x, y, z, and root_r.
Is there any way to resolve this problem?
Thank you.

Environments

• Python 2.7.12

Hi, Thanks for you great work! I've been looking forward to the release of " From human 3D pose estimates to 3D characters" for some interesting applications. Could you please give an approximate estimate of when it will be released?

The number of bone points is the same, but the proportion is not the same. I can directly copy the displacement and rotation vector of character a to the corresponding node of character B. in this way, the same effect can be achieved without using deep learning. Do you think my understanding is correct? @jimeiyang

Thanks!

would you push a pretrained model? thanks!

Hi, the autoencoder loss and cycle loss described in the paper should has a global optimal solution that is just copying quaternions and velocity, but why the GRU model learn something different?

Could I just change the input bvh files in training to retarget a new character?

Will you release the code of demo, whose input is 3D human pose?

Thanks for your awesome work cvpr2018nkn! I have some questions about the visualization code, and your kind help is very much appreciated.

First, I realized that you create a separate .blend file for each motion, and I wonder what’s the difference between two .blend files with same from/to characters. Second, we hope to add more characters for future use, could you please share the script for creating those .blend files? Thank you very much!

Hi, I refer to the paper to make training dataset. However, I fail to find character Malcolm in Mixamo. Would you like to share this data? Thank you very much!!