Faster Human-Machine Collaboration Bounding Box Annotation Framework Based on Active Learning

The conventional method of annotating large object detection datasets is highly time-consuming. To accelerate human-machine collaboration bounding box annotation, a visual framework with human in the loop based on active learning is proposed. Different strategies to prioritize manually annotated images are studied, compared and optimized. Our framework is evaluated on CityPersons dataset. Compared with other stateof-the-art bounding box annotation methods, our proposed approach decreases the total annotation workload by 6.2%.