Semantic Foggy Scene Understanding with Synthetic Data

摘要: This work addresses the problem of semantic foggy scene understanding (SFSU). Although extensive research has been performed on image dehazing and with clear-weather images, little attention paid to SFSU. Due difficulty collecting annotating we choose generate synthetic fog real images that depict outdoor scenes, then leverage these partially data for SFSU by employing state-of-the-art convolutional neural networks (CNN). In particular, a complete pipeline add real, using incomplete depth information is developed. We apply our synthesis Cityscapes dataset Foggy 20550 images. tackled in two ways: 1) typical supervised learning, 2) novel type semi-supervised which combines an unsupervised supervision transfer from their counterparts. addition, carefully study usefulness For evaluation, present Driving, 101 real-world depicting driving come ground truth annotations segmentation object detection. Extensive experiments show learning significantly improves performance CNN Driving; strategy further performance; 3) marginally advances strategy. The datasets, models code are made publicly available.