Semantic Segmentation under Severe Imaging Conditions

Many challenges face semantic understanding of urban street scenes. Two of the most important challenges are foggy and blurred scenes. In this work we make a comparison between two of the most powerful methods in semantic segmentation. These techniques are DeepLabv3+ and PSPNet which achieve the highest mIoU and approximately close to each other on the Cityscapes dataset testing using both fine and coarse data for training. DeebLabv3+ and PSPNet achieved an accuracy of 82.1 % and 81.2% on Cityscapes test set respectively. Our experimental results discuss the performance of these methods on two of the hardest challenges in semantic segmentation which are foggy and blurred scenes.