Structure-Aware Adaptive Diffusion for Video Saliency Detection

This paper proposes a novel saliency model that reveals the long-term info to boost detection accuracy. The saliency estimation of conventional methods heavily depends on the locally revealed short-term info, and they could easily be trapped into imperfect configurations. In contrast, our method can take full consideration of common consistency of those reliable low-level predictions from the perspective of the entire video sequence. Meanwhile, we adopt a newly designed self-learning strategy which is guided by the low-rank analysis to adaptively reveal the long-term spatial-temporal video coherency. To avoid the error accumulations, we also propose a novel non-local descriptor to enhance the discriminative power of the feature space. Thus, the newly revealed the long-term info can be directly regarded as a trustful indicator to sustain additional low-rank analysis, which would serve as the basis toward selective fusion and significantly enhance the detection accuracy.