A State-based Game Attention Model for Cloud Gaming

One of the main promises of cloud gaming, an emerging and growing market in the gaming industry, is its lack of dependence on high-end hardware. To fulfill its goal of enabling anyone to play their favorite games whenever, wherever and on any device, it requires high bandwidth, which remains a major challenge. One solution is to model or predict the players' visual attention map and allocate bitrate accordingly, thereby reducing the bandwidth. The first step of this solution is to predict the players' visual attention maps, which is the objective of our work. In this paper, we demonstrate experimentally that the predicted visual attention maps can be further improved by incorporating game state. Furthermore, we propose a game attention model based on game states. To evaluate the model, we have prepared a 92 minute dataset of states of three games. The results indicate that incorporating game states into visual attention models improves the accuracy of the predicted attention maps by 17.4% on average.