Converting dashcam video into realistic autonomous vehicle sensor data

Training self-driving cars requires diverse, high-quality multi-sensor data (cameras, LiDAR), but AV fleets have limited scale and geographic coverage. This work generates realistic multi-view camera images and LiDAR point clouds from single dashcam videos using diffusion models, trained on synthetic paired data created by rendering real AV logs as dashcam-style videos. The method converts challenging in-the-wild footage into structured formats compatible with AV systems, potentially opening vast internet video sources for training.