NightNext: Nighttime Semantic Segmentation with Decoder-Stage Reconstruction

Abstract Nighttime road-scene parsing is a core function of autonomous driving perception. Low illumination, local overexposure, and motion blur severely degrade boundary clarity and impair the parsing of small objects, elongated structures, and complex semantic transition regions. Most existing methods rely on image enhancement or domain adaptation, but rarely examine the decoding stage where these degradations accumulate. We propose NightNext, a segmentation framework that addresses these challenges through targeted reconstruction of the decoding stage without relying on image enhancement or domain adaptation. We introduce two complementary modules and one training strategy: a Structure-aware Detail Enhancement Module that gates noisy shallow features and refines them through frequency-domain recalibration and multi-scale convolution, a Boundary-aware Detail-Context Fusion Module that adaptively balances detail and semantics using boundary priors, and a Decoupled Boundary Supervision strategy that prevents unstable boundary responses from directly modulating detail features. Experiments on NightCity and ACDC Night show that NightNext delivers strong segmentation accuracy, reaching 57.74 mIoU on NightCity and 55.38 mIoU on ACDC Night while keeping the model at 14.83M parameters.