Creation of Day-Night Visible-Thermal Paired Image Dataset via Image Registration

Michiya Kibe, Takeru Inoue, Junya Morioka, Ryusuke Miyamoto

Abstract: Thermal imaging is challenging because of the lack of color and contrast information. Therefore, translating thermal images to visible color images is crucial for human scene interpretation. Recent advancements in deep learning have significantly improved image translation tasks. Nonethe-less, RGB-Thermal (RGB-T) paired image datasets, particularly for nighttime images, remain scarce. In this study, we propose a method for creating RGB-T paired images by aligning daytime RGB color images with nighttime thermal images at the pixel level by using an image registration technique. Our approach combines the robust RoMa matching method, which effectively handles environmental changes, with simple preprocessing to address modality differences, thus enabling high-precision matching between different modalities. We validate the effectiveness of this method through experiments using the Caltech Aerial RGB-Thermal Dataset captured from drones. As a result, our approach achieves a high matching score of over 95% and facilitates the conversion of nighttime thermal images to interpretable visible images.