Evaluation of the bounding box uncertainty of deep-learning object detection in HALCON software

Abstract

Deep neural networks have become more and more relevant for vision systems, for object detection and classification in industrial fields, such as robot navigation, monitoring and tracking. For such applications, vision systems have to be robust to environment conditions, occlusions and very accurate, as for bin picking. In this paper, we evaluate the performances of deep learning object detection neural networks in HALCON software, by investigating the uncertainty of bounding box position for object detection and the impact of disturbances. In this study, results evidenced the increase of bounding box uncertainty and the reduction of confidence of neural networks when disturbances are introduced, as well as the increment of uncertainty, when confidence lowers. When errors are introduced in labeling, the uncertainty of the bounding box position becomes higher, but lower than the error introduced.