부산대학교 도서관

The development of deep learning models for intelligent vehicles rely on a large number of reliable data, among which large-scale and accurately labeled traffic scene image data is conducive to promoting the research of intelligent perception algorithms. In contrast to the time-consuming and laborious process of manual annotation, the virtual dataset synthesis method is a convenient and efficient way to replace manual labeling. However, due to the existence of automatically labeled samples in the virtual dataset that are not conducive to model training, and the domain shift between the virtual dataset and the real scene, the model trained from the virtual dataset cannot be well generalized to the real scene. To address these issues, we propose a two-stage label screening strategy for the data processing to remove the interference samples with low information content and further propose a novel efficient traffic scene dataset named Large-Scale Software-Defined Transportation Virtual Dataset (LSTVD). Moreover, an attention-induced style alignment domain adaptation method (ASADA) is proposed to reduce the distribution difference between the virtual data domain and the real-world data domain. Experiment results demonstrate the efficiency of our label screening strategy and our domain adaptation method.