Refining Semantic Granularity of Aerial Image Segmentation Datasets Based on Ground Information
Adrian Bauer, Jan-Christoph Krabbe, Mohaned Ibrahim, Anton Kummert · 2023
This paper presents a novel approach for refining the semantic granularity of remote sensing image segmentation datasets using openly available cadastral data, thus minimizing the annotation effort. A two-step method is proposed: (1) a supervised learning step for training a semantic segmentation model on fine-grained cadastre labels to extract class prototypes, and (2) an unsupervised learning step utilizing hierarchical clustering on the extracted prototypes to generate a class hierarchy. The method is demonstrated on building segmentation in aerial imagery data, resulting in models capable of predicting new aggregated semantic classes without extra effort in dataset annotation. Evaluation of the proposed method shows its ability to generate meaningful hierarchical relationships among labels and achieve high segmentation performance.