Abstract
Video saliency detection is typically performed by combining temporal saliency and spatial saliency, which are detected separately. Among various available techniques, uncertainty-based combination is a unique and promising approach. In this paper, we study the uncertainty of each point within a map generated from video saliency detection. We develop an adaptive temporal correlation-based method for unsupervised uncertainty estimation. To evaluate the performance of our algorithm, we propose a systematic evaluation scheme that involves both the creation of ground truth uncertainty data and the comparison of uncertainty estimation results against the ground truth. Our experiments on the public CRCNS database show that the unsupervised uncertainty analysis algorithm is very promising.