Lee Hongjae ; Han Changwoo ; Yoo Jun-Sang ; Jung Seung-Won

Abstract

Semantic segmentation for autonomous driving should be robust against variousin-the-wild environments. Nighttime semantic segmentation is especiallychallenging due to a lack of annotated nighttime images and a large domain gapfrom daytime images with sufficient annotation. In this paper, we propose anovel GPS-based training framework for nighttime semantic segmentation. GivenGPS-aligned pairs of daytime and nighttime images, we perform cross-domaincorrespondence matching to obtain pixel-level pseudo supervision. Moreover, weconduct flow estimation between daytime video frames and apply GPS-basedscaling to acquire another pixel-level pseudo supervision. Using these pseudosupervisions with a confidence map, we train a nighttime semantic segmentationnetwork without any annotation from nighttime images. Experimental resultsdemonstrate the effectiveness of the proposed method on several nighttimesemantic segmentation datasets. Our source code is available athttps://github.com/jimmy9704/GPS-GLASS.