Detection of unexpected multi-part objects from segmented contour maps

A novel method is proposed to detect multi-part objects of unknown specific shape and appearance in natural images. It consists in first extracting a strictly over-segmented map of circular arcs and straight-line segments from an edge map. Each obtained constant-curvature contour primitive has an unknown origin which may be the external boundary of an interesting object, the textured or marked region enclosed by that boundary, or the external background region. The following processing steps identify, in a systematic yet efficient way, which groups of ordered contour primitives form a complete boundary of proper multi-part shape. Multiple detections are ranked with the top boundaries best satisfying a combination of global shape grouping criteria. Experimental results confirm the unique potential of the method to identify, in images of variable complexity, actual boundaries of multi-part objects as diverse as an airplane, a stool, a bicycle, a fish, and a toy truck.