A novel contour detector that refines a rough boundary between an object and a background to a precise boundary in moving pictures robustly is proposed. To estimate boundaries of objects, the proposed method uses self-similar block matching (SSBM) in spatio-temporal 3-D space. SSBM, which searches a larger similar block for each block placed near a boundary, estimates contours correctly. In this paper, it is shown analytically that the robustness of spatiotemporal SSBM is superior to that of conventional 2-D SSBM. Since SSBM does not assume contour smoothness, the proposed algorithm can detect sharp corners more accurately than the methods using smooth constraints such as Snake.

Experimental results show that the proposed method is effective for estimating precise regions of objects even if pictures are noisy.