A Novel Prim's Algorithm Merged with Edge Feature for Fast Object Localization

This paper addresses the low-efficiency of generating possible object locations for use in object detection. In consideration of the existing deficiency of sliding window detection using an exhaustive method, we propose a method of improved Prim's algorithm which combines super pixels and generates proposals windows. The method employs a structured edge detection operator to fast gauging the edge structure, and through the non-maximum suppression engender each edge group, then those groups build object contours to determine the initial locations of random Prim's algorithm combining each super pixel inside, at the end, frontier magnitude between super pixels with the edge initialization can determine the algorithm's termination. Experiments on the challenging PASCAL VOC 2007 dataset show that improved Prim's algorithm's average best overlap has achieved 0.78, the number of proposal windows which are generated by original algorithm is reduced about 60%, while the algorithm's speed is also increased.