Abstract
This paper proposes a video saliency detection model for MPEG and HEVC coded videos. The model extracts features from MPEG macro blocks and HEVC coding units. The feature variables are based on syntax elements and statistics of prediction error. The suitability of the selected features is verified through the use of stepwise regression. Three saliency maps are generated based on intra-frame distances, inter-frame distances and global distances. The proposed model is tested using the eye-1 dataset compiled by Laurent Itti lab in the University of Southern California. The accuracy of the model is quantified by comparing saliency values at human saccade locations against saliency values at random locations. The comparison is performed in terms of Kullback-Leibler distances and receiver-operator curves. The proposed solution is compared against existing work using similar experimental setup. Experimental results revealed that a Kullback-Leibler distance of 2.14 and area under the receiver-operator curve of 0.936 are achieved.
