Automatic Foreground Seeds Discovery for Robust Video Saliency Detection

In this paper, we propose a novel algorithm for saliency object detection in unconstrained videos. Even though various methods have been proposed to solve this task, video saliency detection is still challenging due to the complication in object discovery as well as the utilization of motion cues. Most of existing methods adopt background prior to detect salient objects. However, they are prone to fail in the case that foreground objects are similar with the background. In this work, we aim to discover robust foreground priors as a complement to background priors so that we can improve the performance. Given an input video, we consider motion and appearance cues separately to generate initial foreground/background seeds. Then, we learn a global object appearance model using the initial seeds and remove unreliable seeds according to foreground likelihood. Finally, the seeds work as queries to rank all the superpixels in images to generate saliency maps. Experimental results on challenging public dataset demonstrate the advantage of our algorithm over state-of-the-art algorithms.