ReSIT: A more Realistic Synthetic Driving Dataset for Multi-Domain Image-to-Image Translation
Keywords: Synthetic Dataset, Generative AI, Diffusion, Image Translation
TL;DR: We introduce ReSIT, a synthetic driving dataset with diverse domains for multi-domain I2I translation and a diffusion model that preserves structural content during complex domain translations.
Abstract: Driving dataset is essential for success of autonomous driving system, yet collecting real-world data under diverse domains such as weather, time, and location is challenging and costly. This difficulty results in real-world driving datasets with restricted data domains. Although synthetic driving datasets have been introduced to address this issue, the diversity of domains they can cover remains limited. In this paper, we present ReSIT, a synthetic driving dataset built using a simulation platform that enables precise control over data collection conditions, resulting in more domains and possible combinations than existing datasets. Comparative analyses demonstrate that our dataset is more realistic than previous datasets. Additionally, we present a text-guided diffusion model tailored for multi-domain image-to-image translation, using an adapter for precise source image feature injection and guidance for effective translation. Experimental results show that our model outperforms existing models in preserving the structural content of source images during domain translation even in complex driving scenes. Our code and dataset will be released with the paper.
Submission Number: 23
Loading