Object detection based on scale-invariant partial shape matching

This paper aims at detecting objects via a partial shape matching in unlabeled real images. As both the scale and consistent fragment extraction are troublesome issues in computer vision, we first extract the corresponding parts of pairs of matching fragments generated by the curvature extreme points in object contours. Then, we establish the scale-calculable shape descriptor to keep that the partial shape matching algorithm is scale and rotation invariant. In detection stage, a weighted voting scheme is used to locate candidate object centers and followed by a refinement process to obtain the precise object boundaries. Experiments on ETHZ shape category database validate that using single model shape without training for each category can match (or exceed) the performance of state-of-the-art object detection algorithms