Super-Resolution Data Augmentation Method on 2D Segmentation of 3D Point Cloud Data

3D deep learning methods are widely used in point cloud semantic segmentation. Semantic segmentation of 2D images obtained from 3D point clouds is also an alternative method. This method allows usage ofseveral techniques for data augmentation and preprocessing which might not be available in 3D and could be advantageous in segmentation of classes without height characteristic such as road classes. However, clarity and data loss might appear in the process of dimensionlality reduction. This could negatively effect the segmentation performance. To address this issue, a Super-Resolution data augmentation approach has been proposed. A simple deep learning model was trained using high resolution satellite images of urban areas. Model outputs were used as a data augmentation technique to train binary segmentation models. Results show that SR data augmentation improves segmentation performance of the road class and it could be a viable option for segmentation of images created from 3D point clouds.