Julia Klint presents her thesis

In building analytics, detecting roof superstructures such as dormers and chimneys from satellite or aerial imagery is a common approach for gaining insights into an area. While significant data is available from these sources, this thesis focuses on utilizing drone imagery, which captures higher-resolution images at lower altitudes but introduces different perspectives. To address these challenges, a preprocessing step—orthorectification—was applied to standardize perspectives, making drone images more comparable to aerial images. Although orthorectification proved valuable when applicable, finding a drone dataset with sufficient overlap was challenging. Drone videos were used to extract images, which were then orthorectified, resulting in a diverse set of orthoimages. Additionally, a synthetic dataset generated from a video game was employed. The thesis explores transfer learning by initially training a model on a larger aerial dataset and then fine-tuning it on drone imagery to improve performance. Experiments compared training strategies: starting with random weights, fine-tuning a general instance segmentation model (YOLO V8-l), and fine-tuning a pre-trained aerial model developed by the collaborating company, credium, which was originally based on the same YOLO V8-l model and further fine-tuned on aerial data. Training from random weights yielded very poor results compared to when transfer learning was applied. Decent results were achieved when the YOLO V8-l model was fine-tuned on the datasets, but even better results were observed when the aerial model was used before fine-tuning. This method shows promise when a suitable drone dataset is available or can be generated. Future work could explore whether direct instance segmentation on raw drone images, without orthorectification, provides comparable results, potentially bypassing the preprocessing step.