To complete or not to complete: Gap completion in real images

Gap Completion is a key step in the process of linking of edges into contours towards generating meaningful boundaries. The likelihood of completing a gap between two edges has been approached either indirectly through studying optimal completion contours such as Elastica and Euler Spiral, or through the statistics of co-occurrence. These studies do not address the issue of finding appropriate candidates for gap completion and do not typically address the role of appearance and contextual contours. The paper's contributions are twofold. First, we introduce a Gap Completion Ground-Truth Dataset (GCGD) which annotates human subjects' completions using a local window around a potential gap when contours only or contours and appearance are presented. This dataset is proposed as a mechanism for evaluating gap completion strategies. Second, we show that a shock-based identification scheme identifies gap candidates effectively. We then introduce a scheme for ranking these candidates based on a likelihood measure. The results compare favorably to CDT-based gap completion when evaluated on the GCGD and on the Berkeley Segmentation Dataset.