What's There in the Dark

원문 초록

Scene Parsing is an important cog for modern autonomousdriving systems. Most of the works in semantic segmenta-tion pertains to day-time scenes with favourable weather andillumination conditions. In this paper, we propose a noveldeep architecture, NiSeNet, that performs semantic segmen-tation of night scenes using a domain mapping approach ofsynthetic to real data. It is a dual-channel network, wherewe designed a Real channel using DeepLabV3+ coupled withan MSE loss to preserve the spatial information. In addition,we used an Adaptive channel reducing the domain gap be-tween synthetic and real night images, which also comple-ments the failures of Real channel output. Apart from thedual channel, we introduced a novel fusion scheme to fuse theoutputs of two channels. In addition to that, we compiled anew dataset Urban Night Driving Dataset (UNDD); it consistsof7125unlabelled day and night images; additionally, it has75night images with pixel-level annotations having classesequivalent to Cityscapes dataset. We evaluated our approachon the Berkley Deep Drive dataset, the challenging Mapil-lary dataset and UNDD dataset to exhibit that the proposedmethod outperforms the state-of-the-art techniques in termsof accuracy and visual quality