Visual Attention Based Model

       2012年遥感计算器视觉国际会议——This paper proposes a novel cascade framework to improve spatiotemporal object extraction algorithms for unconstrained image sequences. The cascade framework successively incorporates the constraints on the size of objects for candidate region prediction, an improved backprojection algorithm for coarse localization, ASIFT feature matching for object markers propagation and a novel interactive region merging method for the exact object contour segmentation. Realworld experiments show the effectiveness of the proposed method in the case of varying viewpoint, changing backgrounds, and similar distractors.