Class-specific object contour detection by iteratively combining context information

In this paper, we propose an approach to class-specific object contour detection. Unlike traditional edge/boundary detection algorithms which are designed to detect characteristic changes in brightness, color, and texture, class-specific object contour detection involves the concept of “object”. It aims to capture the main structure (outline) of the object of interest and suppress the strong edge/boundary responses in the background and interior of the object at the same time. Towards this end, we formulate class-specific object contour detection as a supervised learning problem and adopt auto-context to mine the context information (structure similarity among object instances) and combine it with the appearance information in an iterative manner. Experiments on the Weizmann Horse dataset demonstrate that the proposed method is effective and efficient.