bit.shadowplay.gan
We present a shadow play system that displays in real-time the extrapolated motion sequences of the still shadow images casted by players. The crux of our system is a sequential version of the recently developed generative adversarial neural networks with a stack of LSTMs as the sequence controller. We train the model with preprocessed image sequences of diverse shadow figures’ motions, and use it in our enhanced shadow play system to automatically animate the shadows as if they were alive.