PDAOD: An Unsupervised Domain Adaptation Method for Orchard Apple

Traditional unsupervised domain adaptation methods are limited in agricultural object detection due to small target sizes and cluttered backgrounds. The PDAOD-Net model is introduced, offering the capability to dynamically distribute attention between localization and classification tasks in a flexible manner. By utilizing point confidence matrix alignment to achieve precise domain adaptation, experiments show that PDAOD-Net achieves superior performance on apple orchard datasets, outperforming existing methods by approximately 5.8% in AP.