Master Thesis by Jannis Kambach, WWU Münster

Considering the steadily growing volume of available satellite images and the increasing importance of deep learning for analyzing them on a larger scale, the main goal of this work is to evaluate weakly superised learning as an approach to reduce the dependence on manually created labels. As shown by Bearman et al., a segmentation model trained on point-lables can outperform a fully supervised model given the same annotation time budget.

• Most methods rely on pseudo-masks to produce accurate segmentation results - generalized methods that measure the likelihood that an image region contains an object
• Early comparison tests of different objectness methods show poor performance on satellite images compared to images from benchmark sets such as PASCAL VOC
• Laradji et al. propose a custom loss function constructed around the watershed-algorithm that enables full segmentation without pseudo-masks and objectness methods
• Their basic method falls short of the fully supervisied model given the same annotation time budget
• Integrating the COB objectness method into the custom loss function improves segmentation performance
• Training a U-Net on the noisy segmentation maps produced by the improved model improves segmentation performance again

• 0_Objectness: testing notebook for evaluating different objectness methods on satellite images and PASCAL VOC
• 1_Preprocessing: breaks the big 1km x 1km images up into smaller tiles and converts coordinate-based labels into pixel-based labels
• 2_Training: main training loop with configuration options for testing different supervision approaches
• 3_Testing: for importing saved model and evaluating them on the test set

The preprocessing script expects a dataset of satellite images consisting of 1km x 1km GeoTIFF files alongside a shapefile for the labels. The main contributions are the custom loss functions in the losses folder.