Video saliency detection using multi-level spatiotemporal orientation

Abstract

Detecting saliency objects in video is a challenging problem. Conventional saliency detection methods for still images do not take consideration of the motion information, which may fail to detect the moving objects in videos. In this paper, we propose a novel method for detecting saliency objects in videos. Motion cues, which are extracted from both image orientations and video orientations, are integrated with the image cues in order to find the moving objects, We extract "compositions" from each frame to reform the potential shape of the salient object. Additionally, we introduce an extended Spatial-temporal Orientation Energy (SOE) model that computes the motion of objects from the whole video rather than the adjacent frames. Experimental results show that our method outperforms most of the saliency detection methods with various evaluation methods and settings.