Markov Decision Process for Video Generation:

Qualitative Results

Tai Chi Sample

Trained on a single video sequence from the Tai Chi dataset (left), the MoCoGAN model (middle) produces a sample with a freezing pattern occuring directly after the training length (K = 16). Our MDP model (right) uses the same architecture for the generator, but generates longer motion.

Human Actions dataset

The looping pattern is inherent to the motion dynamics present in the Human Actions dataset, hence represents a plausible movement. Here, we show that the video samples produced by our MDP model are visually comparable to video sequences from the MoCoGAN baseline.

UCF-101 dataset

On the more challenging UCF-101 dataset, we show that the videos from the MoCoGAN model (left) consistently exhibit a looping pattern with an approximate period of 24 frames. Samples from the ground-truth dataset (bottom) are looping-free. In comparison, our MDP-based model (right) suffers less from looping. Although we can still recognize the looping pattern for some samples (red boxes), the MDP approach does not exhibit looping artifacts for many other cases (blue boxes).