【摘要】 为了在复杂的场景中均匀突出时空一致的显著物体，提出了一种无监督的视频显著性检测框架。该框架包含3种显著图，运动显著图、空间显著图和时间显著图。首先，本文提出利用运动对比度来突出视频中显著运动的物体，并自适应融合时空梯度场图生成运动显著图；接着，利用帧内的局部对比度和全局对比度对运动显著图中的前景目标进行精细化，得到空间显著图；其次，根据视频序列相邻帧之间的目标存在时间一致性的特点，利用相邻帧之间的颜色和运动相似度加权融合前一帧的显著图，生成时间显著图；最后，融合3种显著图得到视频显著图。本文在3个公开的数据集（DAVIS、FBMS和Segtrackv2）上对本算法进行了定性和定量的评估。实验表明，本算法在具有挑战性的场景中对目标物体的检测性能优于目前经典的显著性检测算法。更多还原

【Abstract】 This paper proposes an unsupervised video saliency detection framework to highlight salient objects with consistent spatio-temporal uniformity in complex scenes. The framework consists of motion saliency maps, spatial saliency maps, and time saliency maps. Firstly, this paper presents the motion contrast to highlight the significant motion of the object, which is adaptively combined with spatio-temporal gradient field map to generate the motion saliency map. The foreground object in the motion saliency map is refined on using the local contrast and global contrast in the frame to obtain a spatial saliency map. Furthermore, according to the characteristics of temporal consistency between consecutive frames, color similarity and motion similarity in two adjacent frames are combined with the saliency map of the previous frame to generate temporal saliency map. Finally, the three saliency maps are transformed into one video saliency map.The proposed method is extensively evaluated on three public datasets(DAVIS, FBMS, and Segtrackv2). Experiments show that the proposed method outperforms the classical saliency detection methods in challenging scenarios.更多还原