Night-to-day thermal image translation for deep thermal place recognition

Place recognition PR is a fundamental task in autonomous robot systems that is still actively being researched. In recent years, CNN-based place recognition methods have surpassed classical methods. However, place recognition in the thermal infrared (TIR) image domain has shown poor performance when applied to both traditional and CNN-based methods due to the appearance variation of the same place throughout the day caused by varying temperature differences. In this paper, we propose a GAN-based nighttime to daytime thermal image translation model that translates thermal images captured at different times of the day into contrast-consistent and detail-preserving images, thus achieving time-agnostic thermal image representations. By applying our GAN-based models to input thermal images for place recognition tasks, we achieved a top-1 accuracy of 80.69% on the STHeReO dataset, outperforming other baseline methods.