nihaomiao / wacv23_tsnet Goto Github PK

@nihaomiao please guide how do you get clean_unseen_video_dict.json and clean_video_dict.json ? I have tried to extract the key-points using openpose but it seems bit different from the json that you have provided under dataset folder.

please say

Thank you for a very good job.
I don't know how to get clean_unseen_video_dict.json and clean_video_dict.json to train the model.
Can you tell me the specific steps or code?

the hand of result is bad, how to improve it ?

Hi, have you tried to fine-tune your pretrained model on a new subject?
How much data is needed for this and how long does it take to fine-tune the model so that it works well on a new subject?

Hi. could you please include something like requirements.txt into your project, so that installation of the necessary libraries would be easier?
I would also like to add that this project seems very promising and a working colab notebook would be greatly appreciated by whole community. ;) Easily testable demo would surely bring in many new fans of this project. Thanks.

Dear author, What method do you use to extract the keypoints of the face in the pose task.

What a wonderful work!
I would like to ask if it is possible to change the subject to anime characters, and then use the driving video of real person dancing to make anime characters dance?
Thank you and look forward to your reply

I see that the label_nc in the train_pose.py is 25. When the model is trained, are there 21 key points of the hand?

how to get clean_video_dict.json and clean_unseen_video_dict.json ?
and what is the function of clean_video_dict.json and clean_unseen_video_dict.json ?
thanks

Hi, thank you for your good project. But there is a question for me. I wanna know where the pose keypoint detector in your codes is. Or is there any tools or model that can transfer the pose keypoint into json file like yours in demo/dance_example/labels/00164/...json.
Thanks!

Currently, I implement training as well as prediction using individual datasets. However, I find that there is still ambiguity in this part of the image hand action. It causes the video to appear uneven and a sense of abruptness can appear.

Hey, when will the pretrained models be released?

Hi, thanks for your great work.
But when I replace your images in "00110" with my own images, and aiming to drive it dance. I have get the pose detection using openpose model and place the json files in the right directory. However, after I execute the demo_pose.py, a Weird female face appeared:

why happened that? Could you please provide some suggestions?

the result of training TS-Net on the new subject is poor ,
LEARNING_RATE = 2e-4
training image num is 296 .

and when the training image is same as testing image .
the result is good,
so it need a large number of images includes different pose to train?

Hi,

Thanks for your work.

The link for the forensicface pretrained model is not working.

Would you be able to update it please?

Hi, I have a few questions regarding the training code it would be great if you could help me.

1. [Full Training] The paper's and code's parameters do not match. which parameters we should use to train fully? we want to replicate your face.pth model which you share on Google Drive. can you please give me those parameters?

2. [Fine Tuning] Can you please also tell me the optimal hyper-parameters for the new subject? I'm training on the parameters mentioned in the code and the output is not that good on the new subject.

Thanks

can you provide the link of dataset of Youtube-dance and FaceForensics ?

why the result lose the hand ? it becomes black mask .