Analysis Of Movement Type Of Natural Dynamic Scenes

Since visual attention-based computer vision applications have gained popularity, ever more complex, biologically inspired models seem to be needed to predict salient locations(or interest points) in naturalistic scenes. Currently, the performance of predictive on baseline models is disappointing. In this paper, we have derived a generic yet powerful model for bottom-up saliency from the simple assumption that the degree of local intensity variation is related to the informativeness of an image region. We use the concept of intrinsic dimension to measure the motion type in natural dynamic scenes. We characterize typical video structures based on the geometrical invariants H, S, and K of the structure tensor, which correspond to the intrinsic dimension of a movie region. Thus, invariants H, S and K of a movie, respectively, indicate the intrinsic dimension of one, two and three. In order to get the feature energy of the video, we need to construct anisotropic spatio-temporal invariant pyramids. Such feature energy vectors are finally fed into a classifier, which learns a mapping between feature energy vectors and the saliency level of a certain location. Despite its conceptual simplicity, our model outperforms state-of-the-art baseline models.