Detecting Keypoints for Automated Annotation of Bounding Boxes using Keypoint Extraction

Abstract

Object detection requires an enormous amount of training data annotated by bounding boxes. All bounding boxes are manually drawn, which leads to highly expensive labor costs. Therefore, this study proposes automatic bounding box annotation of training data for object detection. The keypoints to identify object regions in pictures are extracted, which can then be used for drawing bounding boxes automatically, thus, reducing manual labor requirements. When our proposed method is used for pictures of road signs, keypoints that identify road sign regions in the pictures are detected; these keypoints are found to be highly accurate for drawing bounding boxes.