search

fabro66 / GAST-Net-3DPoseEstimation

A Graph Attention Spatio-temporal Convolutional Networks for 3D Human Pose Estimation in Video (GAST-Net)
MIT License
312 stars 70 forks source link

Reconstruct 3D poses can‘t move like input video #43

Open wtnan2003 opened 3 years ago

wtnan2003 commented 3 years ago

https://user-images.githubusercontent.com/20351890/120980971-bda46480-c7a9-11eb-8a7b-74893c6ff153.mp4

what should I do to fix it?

Any help would be appreciated

hujb48 commented 2 years ago

Hi, @wtnan2003, I also have the same issue, and finally I found that there is the same situation with us in #9. Here, the author preferred to use the existing prior knowledge to design a 3D trajectory model, and you could add: prediction[0] = revise_skes_real_time(prediction[0], re_kpts[0], width) in gen_skes.py with the line 132-138 just like:

   # Adding absolute distance to 3D poses and rebase the height
    if num_person == 2:
        prediction = revise_skes(prediction, re_kpts, valid_frames)
    elif ab_dis:
        # Calculate the trajectory
        prediction[0] = revise_skes_real_time(prediction[0], re_kpts[0], width)
        prediction[0][:, :, 2] -= np.expand_dims(np.amin(prediction[0][:, :, 2], axis=1), axis=1).repeat([17], axis=1)
    else:
        # Calculate the trajectory
        prediction[0] = revise_skes_real_time(prediction[0], re_kpts[0], width)
        prediction[0][:, :, 2] -= np.amin(prediction[0][:, :, 2])

Hope, It can be useful. Yours, Hu.

wtnan2003 commented 2 years ago

@hujb48 Thanks!! I will have a try

JACKYLUO1991 commented 2 years ago

@hujb48 sorry, but what is the params ratio_2d_3d? should set it to... ?

hujb48 commented 2 years ago

@JACKYLUO1991 Sorry, I also have no idea on this parameter, maybe just for the suitable exhibition in the coordinate system, I guess, hope you can let me know if you get the answer.