Dubai Aerial photo segmentation with Unet++! (ResNet101 backbone) The data set is Semantic segmentation of aerial imagery from Kaggle. Was built in Google Colab environment, so make any adjustments needed for it to work on your machine. You can find code and detailed project analysis in aerial_semantic_segmentation.ipynb notebook. The results of the training will be presented below. Dataset consists of 72 unique Dubai Aerial photos and corresponding ground truth masks consisting of 6 classes: 1.Building 2.Land (unpaved area) 3.Road 4.Vegetation 5.Water 6.Unlabeled Model trained for around 250 epocs (2.5 hours) with a batch size of 28 and fine tuned for couple more hours with a batch size of 56. Resulting training metrics were: Below you can see examples of model performance: