Semi-supervised active learning for 3D object detection using consistency based uncertainty

Abstract

With the development of autonomous driving, the perception model in autonomous driving is expected to work safely in various environments. Perception model that understands the surrounding situation of autonomous vehicles is a deep learning model that learns from the labeled sensor data obtained while driving. However, reaching global scalability requires a countless number of data, and labeling the collected data is not only costly but also inefficient due to redundant data. Accordingly, active learning research is actively underway to increase data efficiency. Active learning is an efficient training method that alternates between training and labeling. At the labeling stage, or selection stage, the trained model actively selects the uncertain data to be labeled for the next training stage. However, when training an object detection model using active learning, the dependence on initial training data is high because the model must first detect the object to calculate the model’s uncertainty for that object. In addition, in the case of 3D object detection compared to 2D object detection, there is a difficulty in estimating the size in the depth direction since there are no points on the rear side of the object. Finally, in the case of an autonomous driving dataset, due to class imbalance between major and minor classes, such as cars and cyclists, the selection is biased towards a specific class in the active learning stage. This study proposes an active learning technique applying consistency-based semi-supervised learning. The proposed method maximizes the detection performance by training the model using unlabeled data in the training stage. The data is selected by further estimating the uncertainty of the box predictions and the class predictions. Increased performance of the model due to semi-supervised learning increased object detection rates, which reduced dependence on initial data. In addition, for datasets with a high class imbalance, we found that selecting data by estimating the uncertainty of the bounding box was more effective in the early active learning stage.