Due to the high data requirements of varied applications artificial data poses as a valuable supporting piece for many data-intensive tasks. To overcome the limitations of manually creating data to train machine learning algorithms for a local wildlife monitoring project a specific solution was proposed. Using a simulation large quantities of visual data are created, and evaluated based on trained deep-learning models. This thesis investigates the usability of artificial datasets for image recognition models, with a focus on aerial detection algorithms. A prototype for a data generation framework is created, which exports annotated datasets, that are used to train object detection models based on the YOLOv8 framework. These models are evaluated on real-world data, showcasing the advantages and disadvantages of the pipeline. The trained models achieved high scores on synthetic data but failed to generalize well on real-world footage. Models trained on both artificial and real data performed reasonably well, which shows the potential of bolstering already existing datasets with matching synthetic data. This provides additional avenues for improving dataset volume without major reductions in quality.
Drone Footage Synthesis Data Generation for Wildlife Monitoring
Kaindl, J. (Author). 2025
Student thesis: Master's Thesis