Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

RHRSegNet: Relighting High-Resolution Night-Time Semantic Segmentation

Sarah Elmahdy, Rodaina Hebishy, Ali Hamdi

TL;DR

This work tackles nighttime semantic segmentation by introducing RHRSegNet, which combines a relighting module with a High-Resolution Network to enhance feature quality under low illumination. The relighting component uses residual convolutional learning and shared weights to simulate lighting changes, feeding enriched features into HRNet for pixel-level segmentation. Across Night-city, Dark Zurich, and Cityscapes datasets, the approach achieves state-of-the-art performance, notably improving mIoU in both cross-domain adaptation (Cityscapes→Dark Zurich) and standard evaluation, and showing strong generalization on NightCity-Fine without domain adaptation. The findings suggest practical significance for autonomous driving and other night-vision applications, offering a robust strategy to mitigate lighting-induced semantic ambiguities in urban scenes.

Abstract

Night time semantic segmentation is a crucial task in computer vision, focusing on accurately classifying and segmenting objects in low-light conditions. Unlike daytime techniques, which often perform worse in nighttime scenes, it is essential for autonomous driving due to insufficient lighting, low illumination, dynamic lighting, shadow effects, and reduced contrast. We propose RHRSegNet, implementing a relighting model over a High-Resolution Network for semantic segmentation. RHRSegNet implements residual convolutional feature learning to handle complex lighting conditions. Our model then feeds the lightened scene feature maps into a high-resolution network for scene segmentation. The network consists of a convolutional producing feature maps with varying resolutions, achieving different levels of resolution through down-sampling and up-sampling. Large nighttime datasets are used for training and evaluation, such as NightCity, City-Scape, and Dark-Zurich datasets. Our proposed model increases the HRnet segmentation performance by 5% in low-light or nighttime images.

RHRSegNet: Relighting High-Resolution Night-Time Semantic Segmentation

TL;DR

This work tackles nighttime semantic segmentation by introducing RHRSegNet, which combines a relighting module with a High-Resolution Network to enhance feature quality under low illumination. The relighting component uses residual convolutional learning and shared weights to simulate lighting changes, feeding enriched features into HRNet for pixel-level segmentation. Across Night-city, Dark Zurich, and Cityscapes datasets, the approach achieves state-of-the-art performance, notably improving mIoU in both cross-domain adaptation (Cityscapes→Dark Zurich) and standard evaluation, and showing strong generalization on NightCity-Fine without domain adaptation. The findings suggest practical significance for autonomous driving and other night-vision applications, offering a robust strategy to mitigate lighting-induced semantic ambiguities in urban scenes.

Abstract

Night time semantic segmentation is a crucial task in computer vision, focusing on accurately classifying and segmenting objects in low-light conditions. Unlike daytime techniques, which often perform worse in nighttime scenes, it is essential for autonomous driving due to insufficient lighting, low illumination, dynamic lighting, shadow effects, and reduced contrast. We propose RHRSegNet, implementing a relighting model over a High-Resolution Network for semantic segmentation. RHRSegNet implements residual convolutional feature learning to handle complex lighting conditions. Our model then feeds the lightened scene feature maps into a high-resolution network for scene segmentation. The network consists of a convolutional producing feature maps with varying resolutions, achieving different levels of resolution through down-sampling and up-sampling. Large nighttime datasets are used for training and evaluation, such as NightCity, City-Scape, and Dark-Zurich datasets. Our proposed model increases the HRnet segmentation performance by 5% in low-light or nighttime images.
Paper Structure (16 sections, 3 equations, 5 figures, 3 tables)

This paper contains 16 sections, 3 equations, 5 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Nighttime Semantic Segmentation
  4. Semantic Segmentation
  5. Research Methodology
  6. Relighting Model
  7. Learning Parameter Sharing
  8. Semantic Segmentation Network
  9. Data Augmentation
  10. Datasets
  11. Night-city
  12. Dark-Zurich
  13. Cityscape
  14. Experimental Design
  15. Results
  16. ...and 1 more sections

Figures (5)

  • Figure 1: Relighting Model
  • Figure 2: Our HRNet Architecture
  • Figure 3: Original Image
  • Figure 4: Ground Truth
  • Figure 5: Our RHRSegNet