A method for improving accuracy of object detection or recognition in an annotation system while minimizing human effort is disclosed. This system runs an already trained detector on unannotated data for detecting objects from data such as image or video. Decisions are availed from human annotation only as the detected object is correct or incorrect. Images confirmed as not containing the object are then fed as ‘negative’ data. Images confirmed as containing the object can be further annotated for the exact bounding box, and then fed as ‘positive’ data to the machine learning algorithm. The process of presenting data to the algorithm can be iterated until the detector accuracy saturates. The advantages of using the method include improved object detection accuracy with minimal human effort.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.