GAN-Based Semantic-Aware Translation for Day-to-Night Images

Perception in autonomous driving has achieved robustness and high accuracy through deep learning. CNN-based methods require a large amount of data collection and annotation. However, most of the current datasets are built on daytime scenes, and there are few datasets for adverse conditions such as night-time. Recently, data augmentation by image-to-image translation using Generative Adversarial Networks (GANs) has attracted attention. GANs based image-to-image translation performs well for various image translation tasks. On the other hand, semantic information may be lost in problems with the significant domain gap, such as day and night. In this paper, we propose a semantic-aware image translation. This framework preserves semantic consistency by transfer learning a semantic segmentation network to GANs. Experimental results show that the proposed method achieved to generate natural night images compared to previous studies.