Visual Object Detection for Drones: A Domain Adaptation Perspective

In recent years, AI applications are being developed for many mobile platforms, including drones. However, domain shifts, i.e., variations in data distributions, often cause performance degradation of deep learning models, when the application is deployed to an environment different from its training data. Few previous works in drone-based object detection have considered the situation when no annotated data will be available for the new working site. Therefore, this thesis proposes a domain adaptation framework for aerial object detection, i.e., DAerialDet, which aims to improve detection accuracy using unsupervised learning methods. With a labeled source domain and an unlabeled target domain, DAerialDet utilizes adversarial domain clsassifiers to align feature distributions of the two, and at the same time enforces temporal consistency across the training time with self-training. Evaluation on CityScapes to Foggy CityScapes adaptation has shown DAerialDet is comparable to recent state-of-the-art. Adaptation experiments with drone-based datasets demonstrate DAerialDet can improve target domain performance with labeled source domain images and unlabeled target domain image sequences