Data Quality Assurance in Autonomous Driving Systems

Abstract

In recent years, autonomous driving systems (ADSs) using deep learning-based modules have significantly attracted the attention of researchers from different communities, such as computer vision. These intelligent systems require a precise and accurate training process before their deployment to real-life situations. The performance and reliability of ADSs are dependent on two important factors, namely, training dataset and model components, each of which must be carefully taken into consideration. Since in most of the realistic cases, the models of ADSs are released in a black-box form, and access to their components (e.g., loss functions and hyper-parameters) is not granted, therefore, ensuring the quality of the samples in the ADSs training datasets is of paramount importance. In view of these explanations, in this work, we focus on developing an efficient scheme for cleaning the training datasets of ADSs that employ deep image object detectors, by identifying the samples in the dataset with erroneous bounding boxes. In this regard, we leverage the visual signals associated with the bounding boxes, in addition to their spatial coordinates, for predicting the erroneous status of the bounding boxes in an accurate manner. Moreover, we incorporate confident learning in the proposed scheme in order to prune the predictions of the erroneous statuses of the bounding boxes, and, further contribute to developing secure and reliable ADSs. The results of the extensive experiments demonstrate the effectiveness of various ideas employed in the design of the proposed erroneous bounding box detection scheme for the ADSs datasets. Further, it is shown that the proposed scheme could significantly outperform the other state-of-the-art data selection methods in cleaning the training datasets of ADSs.