Multi-Label Image Segmentation Via Point-Wise Repetition

Gang Zeng,Luc Van Gool

DOI: https://doi.org/10.1109/CVPR.2008.4587418

2008-01-01

Abstract:Bottom-up segmentation tends to rely on local features. Yet, many natural and man-made objects contain repeating elements. Such structural and more spread-out features are important cues for segmentation but are more difficult to exploit. The difficulty also comes from the fact that repetition need not be perfect, and will actually rather be partial, approximate, or both in most cases. This paper presents a multi-label image segmentation algorithm that processes a single input image and efficiently discovers and exploits repeating elements without any prior knowledge about their shape, color or structure. The algorithm spells out the interplay between segmentation and repetition detection.The key of our approach is a novel, point-wise concept of repetition. This is defined by point-wise mutual information and locally compares certain neighborhoods to accumulate evidence. This point-wise repetition measure naturally handles imperfect repetitions, and the parts with inconsistent appearances are recognized and assigned with low scores. An energy functional is proposed to include the point-wise repetition into the image segmentation process, which takes the form of a graph-cut minimization. Real scene images demonstrate the ability of our algorithm to handle partial and approximate repetition.