Continuous learning with random memory for object detection in robotic applications

If the robot is to interact with environment designed for humans it has to be able to cope with new objects in it’s surrounding, and not only to classify but also effectively localize objects in its reach based on visual sensing. We cannot predict all objects the robot will face. So it needs to learn them while operating, without an external update of software. One of the promising approaches to handle this task is using detectors based on convolution neural networks (CNN) with continuous learning setup. However, there is a well-known problem of catastrophic forgetting (CF) that negatively affects accuracy of neural networks, the main algorithms in object recognition nowadays. Moreover, we need to reach a trade off between training time (because it is unacceptable if the robot trains for hours) and good acccuracy of detection. In this paper we review some of the latest approaches in the field and adapt one of SOtA methods in classification task to detection task. Particularly, we adapt YOLOv5 to work in scenario with continuous learning by adding random memory mechanism from [1] and applying layer freezing in order to increase training speed. We benchmark the proposed method using iCubWorld Transformations dataset and report results with significantly improved metrics and inference speed.