High resolution instance segmentation for building blueprint vectorization

This thesis proposes a high-resolution instance segmentation method based on metric learning approaches for floorplan images with intricate details called blueprints. Our approach first divides an input blueprint image into an overlapping array of crops. Second, we use a metric-learning based instance segmentation technique followed by a clustering algorithm to extract instances. Finally, the segmentation results from overlapping crops are merged using boundary extraction. This approach is simple and achieves performance that is both qualitatively and quantitatively more accurate than the competing methods by a large margin.