Benchmark evaluation of unsupervised domain adaptive object detectors in foggy UAV environments
摘要
Object detection models typically rely on large labeled datasets to achieve robust generalization. However, in practical applications, labeled data from the target domain is often limited or nonexistent, causing supervised methods to struggle when training and testing data come from different distributions. Unsupervised domain adaptive methods overcome this limitation by enabling knowledge transfer from a fully labeled source domain to an entirely unlabeled target domain, making it particularly useful in real-world applications where labeled data is scarce. Despite recent progress, existing benchmarks primarily focus on ground-level imagery (e.g., Cityscapes) or clear-weather aerial views, leaving a significant gap in evaluating robustness under adverse aerial conditions. This paper presents a novel benchmark for unsupervised domain adaptation object detection, based on two UAV datasets: UIT-Drone and UIT-Drone-Foggy, a synthetic fog dataset derived from UIT-Drone. Distinguishing our work from prior studies, we fully reproduce and evaluate ten state-of-the-art unsupervised domain adaptive object detection techniques using their official implementations, training them entirely from scratch without reliance on pre-trained models, revealing gaps in performance when applied to real-world data, and demonstrate the effectiveness of the proposed benchmark for advancing unsupervised domain adaptation research in object detection. To assess the high performance observed on our benchmark in a real-world context, three representative methods, namely DATR, DA2OD, and SIGMA++, which demonstrate high performance on the UIT-Drone