We present a novel object detection strategy using depth cue which is robust to small-sized objects. In practical cases, the target objects that we want to detect often appear in very small portion of the input images and most of previous object recognition and detection ap-proaches aiming to achieve high accuracy for several object databases are not applicable. Adopt-ing the depth cue as a prior removes the scale ambiguity and allows us to use adaptive scales for all candidate regions even when the object is seen at small size. The depth cue can be easily ob-tained by a Laser Range Finder (LRF) or a stereo camera which have become common for many camera-based configurations. By using our proposed method in combination with any other object detection/recognition method, we can expect a substantial improvement of success rate when an insufficient appearance of object causes trouble. We designed an experiment with a scenario in which an indoor mobile robot tries to find several target objects in general environment. The expe-rimental result demonstrates that the proposed method is effective to improve the success rate of small object detection tasks.
조회 수 2230 댓글 0
|Jaesik Park, Yekeun Jeong, Chaehoon Park, In So Kweon
|16th Korea-Japan Joint Workshop on Frontiers of Computer Vision(FCV)