A1520
Title: Enhancing nighttime road image segmentation with CycleGAN-based image transformation
Authors: Jiho Lee - South Korea, Inha University (Korea, South) [presenting]
Donghyeon Yu - Inha University (Korea, South)
Abstract: Image segmentation plays an important role in image recognition in computer vision, as it provides not only object labels but also the boundaries of the objects. However, annotated datasets are insufficient compared to other image recognition tasks, such as object detection with bounding boxes, since datasets for image segmentation require pixel-level object labels. In particular, it is more difficult to annotate pixel-level object labels for images obtained at night. In addition, the performance of image recognition decreases for nighttime images, which are affected by noise from car headlights and streetlamps. In this study, we propose a novel framework that integrates cycleGAN-based image transformation with state-of-the-art segmentation models. Experiments on road images demonstrate that our approach significantly enhances segmentation performance for nighttime images.
