MIPE: Towards Cleaning the Datasets of Autonomous Driving Systems using Multi-modal Information Processing and Confident Learning

Numerous autonomous driving systems employ deep learning-based image object detection schemes for their navigation. For developing reliable autonomous driving systems, the training process of the deep image object detectors must be performed in a precise manner. Existence of samples with erroneous labels, e.g., erroneous bounding boxes, in the training datasets of autonomous driving systems leads to a reduction in their performance and a decrease in their reliability in real-life situations. Given these explanations, in this paper, we propose a novel erroneous bounding box detection scheme for identifying the bounding boxes in the training datasets that are annotated wrongly, and avoid their use in the training process of autonomous driving systems. Specifically, we employ two efficient techniques, namely, multi-modal information processing and confident learning, in the development of the proposed scheme. In the multi-modal information processing, we first obtain the instance segmentation maps of the images using deep image instance segmentation networks, and then, utilize their interactions with the spatial coordinates of the bounding boxes, along with the spatial coordinates themselves, to generate discriminative sets of features for the task of erroneous bounding box detection. Further, by using the confident learning technique, we leverage the statistical information of the estimated erroneous statuses of the bounding boxes, to further enhance the performance of the task of erroneous bounding box detection. The results of extensive experimentations demonstrate the effectiveness of the proposed erroneous bounding box detection scheme in cleaning the datasets of autonomous driving systems, compared to the state-of-the-art data selection schemes.