A Weakly Supervised Strategy for Learning Object Detection on a Humanoid Robot

摘要

Research in Computer Vision and Deep Learning has recently proposed numerous effective techniques for detecting objects in an image. In general, these employ deep Convolutional Neural Networks trained end-to-end on large datasets annotated with object labels and 2D bounding boxes. These methods provide remarkable performance, but are particularly expensive in terms of training data and supervision. Hence, modern object detection algorithms are difficult to be deployed in robotic applications that require on-line learning. In this paper, we propose a weakly supervised strategy for training an object detector in this scenario. The main idea is to let the robot iteratively grow a training set by combining autonomously annotated examples, with others that are requested for human supervision. We evaluate our method on two experiments with data acquired from the iCub and R1 humanoid platforms, showing that it significantly reduces the number of human annotations required, without compromising performance. We also show the effectiveness of this approach when adapting the detector to a new setting.