Shape detector for generic ball detection
Yuncheng Li, Yukun Zhu, Rui Zhang, Jun Wei Zhou · 2015
This paper describes a novel approach for generating object proposals for ball detection. Our method, called shape detector, captures the possible contours of balls and then transfers them into proposal bounding boxes which may contain the target object. These proposal bounding boxes can be further used in class-specific object detection task. Our experiment results on part of ILSVRC dataset show that shape detector can achieve 71.13% recall and a mean average best overlap of 0.648 using less than 300 proposals. It also shows strong performance in object detection, in which we get a mean average precision of 34.33%.