Parking Lot Instance Segmentation from Satellite Imagery through Associative Embeddings

In this paper, we apply the technique of instance segmentation through associative image embeddings using stacked hourglass networks to the problem of parking lot instance segmentation in satellite imagery. We sought an instance segmentation method that, unlike other common methods such as Mask R-CNN, was independent of object classification and robust to missing labeled instances. We demonstrate how associative image embeddings can be created, which can provide instance segmentation for use with any semantic segmentation map.