How does human interest modeling help in computer vision: Tracking-by-saliency in unconstrained social videos

Sample quality plays an important role in tracking-by-learning strategies, but the reliable online samples are hard to be obtained due to challenges of variational environments. By modeling how human visual interest actively guiding the seek of salient regions and movements in video sequences, in this paper, a compositional tracking strategy is proposed based on an integrated saliency map, which is able to accurately guide the process of online samples generation. Meanwhile, a segmentation based refinement method is also proposed for effective model updating. With a high performance kernelized correlation filter, the proposed tracking can efficiently handle the complex intrinsic and extrinsic appearance changes. Experiments on challenging benchmark databases demonstrate that the robust accuracy of the proposed tracking against with the other state-of-the-art trackers.