Unsupervised Video Prediction Network with Spatio-temporal Deep Features

Predicting the future states of things is an important performance form of intelligence and it is also of vital importance in real-time systems such as autonomous cars and robotics. This paper aims to tackle a video prediction task. Previous methods for future frame prediction are always subject to restrictions from environment, leading to poor accuracy and blurry prediction details. In this work, we present an unsupervised video prediction framework which iteratively anticipates the raw RGB pixel values in future video frames. Extensive experiments are implemented on advanced datasets — KTH and KITTI. The results demonstrate that our method achieves a good performance.