Interactive object annotation based on one-click guidance

Due to the large workload of manual annotation of datasets, uneven data quality and high professional thresholds have been a problem. Based on the idea of semi-automatic annotation, this article discusses the method use of interactive methods to obtain accurate annotations of objects. We propose a method of human–machine interactive object annotation based on one-click guidance. Specifically, we click on a point close to the center of the object and use the prior information of this point to give a guide to the model. The advantages of our method are fourfold: (1) the simulated click method is transferable and can be labeled across datasets; (2) clicks help to eliminate irrelevant areas within the bounding box; (3) the operation is more convenient and does not require artificial boxes, we only need to give the relevant location information; (4) our method supports additional click annotations for further correction. To verify the effectiveness of the proposed method, we conducted a lot of experiments on the KITTI and PASCAL VOC2012 datasets, and the results proved that our model has improved average IoU by 18.1% and 14.6% compared with Anno-Mage and CVAT, respectively. Our method focuses on improving the accuracy and efficiency of annotation, and provides a new idea for the field of semi-automatic annotation.