Video Frame Prediction Using Convolutional LSTM Networks

Abstract

This thesis evaluated Convoultional LSTM (ConvLSTM) for frame prediction to help better understand motion in neural networks. Three different neural networks were implemented and trained. The three networks included, the original ConvLSTM paper by Shi et al. [35], the Spatio-Temporal network by Patraucean et al. [32], and the VPN by Kalchbrenner et al. [21]. Implementing and training the three networks allowed thorough testing and comparison of the accuracy and performance of the networks. The original ConvLSTM was reasonably easy to re-implement and train, and it performed as expected. The Spatio-Temporal network and VPN were unable to achieve the desired results, but still showed improvements over the ConvLSTM. The VPN outperformed both the ConvLSTM and spatio-temporal networks and performed better than the reported results in Patraucean et al. [32]. The VPN also showed that it over-fits the data.