Abstract
High-quality, large-scale simulators are pivotal in conducting system testing for developing autonomous driving systems, ensuring safety and efficiency. Traditional simulators, typically based on game engines, necessitate the creation of three-dimensional environments that mimic real-world driving scenarios. Nevertheless, creating and maintaining these digital environments incur significant expenses associated with reflecting evolving real-world conditions. Furthermore, the gap arises in physical properties, such as geometric shapes and optical properties, compared to the real world. To address these limitations, recent research has focused on developing neural simulators, leveraging Neural Radiance Fields (NeRF), a deep learning model facilitating view synthesis, to construct simulation environments. However, the assessment of neural simulators is currently constrained to the validation of reconstruction accuracy alone, necessitating additional validation efforts to explore their applicability within real-world autonomous driving systems. This study presents a comprehensive analysis encompassing the data acquisition methodology, reconstruction accuracy assessments, and evaluation of practical applications in autonomous driving tasks.
