| "Synthetic data has become an essential component of machine learning-based perception in the field of autonomous driving. However, it cannot completely replace real data due to the sim-to-real domain shift. In this work, we propose a method that combines data augmentation and adversarial training to synthesize realistic data for pedestrian recognition tasks. Our approach uses an attention mechanism guided by an adversarial loss to learn and improve the domain discrepancies in sim-to-real adaptation. Our experiments show that our proposed method is robust to such discrepancies and produces visually realistic and semantically consistent data. Additionally, we evaluate our data generation pipeline on pedestrian recognition tasks and demonstrate that the generated data resemble properties of real-world data." |