Abstract
We introduce DyNFL, a novel neural field-based approach for high-fidelityre-simulation of LiDAR scans in dynamic driving scenes. DyNFL processes LiDARmeasurements from dynamic environments, accompanied by bounding boxes of movingobjects, to construct an editable neural field. This field, comprisingseparately reconstructed static backgrounds and dynamic objects, allows usersto modify viewpoints, adjust object positions, and seamlessly add or removeobjects in the re-simulated scene. A key innovation of our method is the neuralfield composition technique, which effectively integrates reconstructed neuralassets from various scenes through a ray drop test, accounting for occlusionsand transparent surfaces. Our evaluation with both synthetic and real-worldenvironments demonstrates that \ShortName substantial improves dynamic scenesimulation based on LiDAR scans, offering a combination of physical fidelityand flexible editing capabilities.