Related Experiment Video
Updated: Jul 2, 2025
03:31
Author Spotlight: Enhancement of Salient Object Detection for Smart Grid Applications
Published on: December 15, 2023
537
Semantic and Geometric-Aware Day-to-Night Image Translation Network
Geonkyu Bang1, Jinho Lee1, Yuki Endo2
1Emerging Design and Informatics Course, Graduate School of Interdisciplinary Information Studies, The University of Tokyo, 4 Chome-6-1 Komaba, Meguro-ku, Tokyo 153-0041, Japan.
Sensors (Basel, Switzerland)
|February 24, 2024
Summary
This study introduces an unsupervised network for day-to-night image translation, improving autonomous driving perception in adverse weather. The method uses semantic and geometric attention maps for better visual results.
Area of Science:
- Computer Vision
- Artificial Intelligence
- Autonomous Systems
Background:
- Autonomous driving systems rely on perception, but current datasets lack adverse conditions like rain and night.
- Existing deep learning models struggle with perception tasks in low-visibility environments.
Purpose of the Study:
- To develop an unsupervised image-to-image translation network for day-to-night conversion.
- To address the challenge of training perception models with unpaired data for adverse conditions.
Main Methods:
- Propose an unsupervised network for day-to-night image translation.
- Extract semantic and geometric information using multi-task learning for semantic segmentation and depth estimation.
- Integrate extracted information as spatial attention maps in an image-to-image translation network.
Main Results:
- The proposed method demonstrates qualitative and quantitative improvements over existing approaches.
- Enhanced visual presentation in day-to-night image translation.
Conclusions:
- The unsupervised network effectively translates day images to night, enhancing perception model robustness.
- The attention map integration approach shows promise for handling unpaired data in domain translation for autonomous driving.
