What: Temporal Autoencoder for Predicting Video

How: Tensorflow version of CNN to LSTM to uCNN

python main.py

sh mergemov.sh

smplayer out_all.mp4

or

smplayer out_all2_fast.mp4

Choose global flags
In main():
- Choose to use checkpoints (if exist) or not: continuetrain
- type of model: modeltype
- number of balls: num_balls

Try more depth in time
Train with geodesic acceleration (can't be done in python in tensorflow)
Try homogenous LSTM/CNN architecture
Include depth in CNN even if not explicitly 3D data, to avoid issues with overlapping pixel space causing diffusion
Estimate velocity field in rgb, to avoid collisions most likely state as averaging to no motion due to L2 error's treatment of two possible states.
Use entropy generation rate to train attention where can best predict.
Try rotation, faces, and ultimately real video.