MoCoGAN: Decomposing Motion and Content for Video Generation

• Powered by DL HACKS

Description

• A pytorch implmention of MoCoGAN
• authors: Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, Jan Kautz
• arxiv: https://arxiv.org/abs/1707.04993
• project page: https://github.com/sergeytulyakov/mocogan

Implemention

• you need to place datasets([run, walk, skip, ......]) here under 'raw_data' dir.
• python resize.py: resize your raw data as 96x96 and place them under resized_data dir.
• python train.py: train MoCoGAN.

Results

Requirements

• python 3.5
• pytorch 0.2
• numpy
• skvideo
• ffmgeg (for preprocessing and saving videos)

References

• L.Gorelick,M.Blank,E.Shechtman,M.Irani,andR.Basri. Actions as space-time shapes. PAMI, 29(12):2247–2253, 2007. http://www.wisdom.weizmann.ac.il/%7Evision/SpaceTimeActions.html (datasets)
• https://github.com/pytorch/examples/tree/master/dcgan