This is the authors' abstract. We don't add key points for this paper.
Generative Artificial Intelligence (AI) is emerging as a transformative tool for synthetic dataset generation in Advanced Driver-Assistance Systems (ADAS). Traditional Autonomous Vehicle (AV) datasets face high collection costs, limited edge-case diversity and scalability challenges, making alternative data generation approaches essential. This paper introduces an innovative application of multi-modal generative AI models for AV perception tasks with a focus on ADAS applications. Specifically, the study explores Text-to-Image (T2I), Text-to-Video (T2V), Image-to-Video (I2V), and Image-to-Image (I2I) models to generate high-fidelity data for object detection, segmentation, and trajectory prediction. Despite their promise, practical deployment is obstructed by prompt adherence failures, mode collapse, uncontrollability, domain adaptation inconsistencies, and computational inefficiencies. We provide a comparative evaluation of these models, benchmarking their realism, consistency and adaptability in bridging the synthetic-to-real gap and enhancing AV robustness. Experimental results highlight the strengths and limitations of different approaches, with insights into controllable data synthesis and improvements in real-time efficiency. Furthermore, YOLO models fine-tuned on generative AI-generated datasets achieve superior detection performance with GAI_YOLOv7 reaching 74.48% mAP and GAI_YOLOv8x achieving 76.16% mAP, outperforming VR-based models (29.76%, 23.44%).
Answers are grounded in this paper and cite their sources.
No takes yet. Share an insight, caveat, or question.
G et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: