Real-time salient object detection engine for high definition videos

Abstract

Salient object detection can provide the attention information of viewers and has the potential to be employed in many real-time interactive and display systems. In this paper, based on an existing top-down salient object detection algorithm, a real-time salient object detection engine is proposed and implemented. With several memory bandwidth optimization techniques and hardware optimization techniques, such as subsampled integral histogram, data reuse, partial sum reuse, and careful system scheduling, an efficient engine is designed. It can support salient object detection for videos at 1920×1080 30fps with reasonable hardware cost.