Robust spatio-temporal saliency estimation method for H.264 compressed videos

This paper presents a robust spatio-temporal saliency estimation method based on modeling motion vectors and transform residuals extracted from the H.264/AVC compressed bitstream. Spatial saliency is estimated by analyzing the detailed sub-band coefficients obtained by the wavelet decomposition of the luminance component of the macro-blocks, while temporal saliency is estimated by modeling the block motion vector orientation information using local derivative patterns. Dempster Shafer fusion rule is used to fuse the spatial saliency map and the motion saliency map to obtain the final saliency for a video frame. Extensive experimental validation along with comparative analysis with state-of-the-art methods is carried out to establish the proposed saliency method.