Uncertainty-Based Semi-Supervised Object Detection in Autonomous Driving Environment

In real-world traffic scenarios, due to the lack of actual labels, it is necessary to introduce a semi-supervised object detection (SSOD) algorithm, which can greatly reduce costs and improve model performance by mining a large amount of unlabeled data and a small amount of labeled data. The improvement of most existing SSOD algorithms relies on the previous semi-supervised classification tasks, ignoring the impact of the regression on the accuracy and robustness of the model, especially in complex autonomous driving environments. Hence, to solve this problem, we propose an uncertainty-based SSOD algorithm framework for the first time, which can quantify uncertainty by self-learning for teacher-student model optimization. And the uncertainty fusion mechanism is used to update the fuzzy labels generated by multiple iterations to improve their quality and ensure the stability of the training process. Finally, exhaustive experiments using state-of-the-art algorithms demonstrate that our proposed algorithm performs best in general object detection settings as well as autonomous driving settings.