Chris Pollett >
Students >
Lei
( Print View)
[Bio]
[Blog]
[GANs-PDF]
[Gen Videos Scene Dynamics-PDF]
[3D CNNs Human Action RECOG-PDF]
[TGAN-PDF]
Implement a video GAN to create fake videos
Description:
This is the first atempt to generate a video with 3D GAN.
Dataset:
UCF101-Action Recognition Data Set and Weizmann Action database
Download a sample of the train videos
Dependencies:
1) numpy
2) keras
3) matplotlib
4) skimage
5) skvideo
3D GAN:
1) Discriminator - 3D Convolutional network.
Plot of the Discriminator Model:
2) Generator -3D Convolutional network.
Plot of the Generator Model:
3) Use Adam optimizer and fixed learning rate 0.0002.
4) Use dropout rate 0.4.
Download 3d_video_gan.py
Output:
128 by 128 mp4 videos with 8 frames.
Download predicted video1
Download predicted video2
Conclusion:
1) 3D GAN alone cannot predict plausible videos
2) Extend training time cannot always create better quality videos
3) The 3D GAN works better for low-resolution videos