FrozenDrive: Zero-Shot Text-Guided Driving Scene Generation and Data Augmentation with Parameter-Free Frozen Diffusion Model
Researchers have developed FrozenDrive, a novel framework for generating synthetic driving scenes using parameter-free diffusion models. This method addresses limitations in current models by preserving pre-trained knowledge and improving text alignment, enabling the creation of consistent multi-view and temporally coherent scenes, even under adverse weather conditions or for rare object classes. When applied to the nuScenes dataset, data augmented with FrozenDrive significantly enhanced the performance and robustness of autonomous driving models, particularly in challenging scenarios like nighttime and rain. AI
IMPACT Enables more robust and diverse training data for autonomous driving systems, potentially accelerating their development and deployment.