STAM: A SpatioTemporal Attention Based Memory for Video Prediction

Video prediction has always been a very challenging problem in video representation learning due to the complexity in spatial structure and temporal variation. However, existing methods mainly predict videos by employing language-based memory structures from the traditional Long Short-Term Memories (LSTMs) or Gated Recurrent Units (GRUs), which may not be powerful enough to model the long-term dependencies in videos, consisting of much more complex spatiotemporal dynamics than sentences. In this paper, we propose a SpatioTemporal Attention based Memory (STAM), which can efficiently improve the long-term spatiotemporal memorizing capacity by incorporating the global spatiotemporal information in videos. In the temporal domain, the proposed STAM aims to observe temporal states from a wider temporal receptive field to capture accurate global motion information. In the spatial domain, the proposed STAM aims to jointly utilize both the high-level semantic spatial state and the low-level texture spatial states to model a more reliable global spatial representation for videos. In particular, the global spatiotemporal information is extracted with the help of an Efficient SpatioTemporal Attention Gate (ESTAG), which can adaptively apply different levels of attention scores to different spatiotemporal states according to their importance. Moreover, the proposed STAM are built with 3D convolutional layers due to their advantages in modeling spatiotemporal dynamics for videos. Experimental results show that the proposed STAM can achieve state-of-the-art performance on widely used datasets by leveraging the proposed spatiotemporal representations for videos.