O-DINO: An Optimized DINO for Accurate Oriented Object Detection in Remote Sensing Imagery

With the emergence of DINO, Transformer-based detection architectures have achieved significant success across various vision tasks. However, despite DINO’s potential performance advantages, its application in oriented object detection remains insufficiently explored, particularly in dense scenarios where current Transformer-based methods underperform compared to CNN detectors.We identify two primary bottlenecks: the original attention mechanism exhibits misalignment with oriented objects, and DINO’s noisy query generation scheme is not well-adapted to oriented object distributions. Hence, we propose O-DINO, an optimized DINO-based oriented object detector. Specifically, we introduce anOriented Deformable Attention (ODAttn)module to align features with arbitrary orientations, and aGaussian Distribution Sampling based Noisy Query Generation (GSNQG)method to better adapt the training process to oriented object geometries. Extensive experiments on three challenging remote sensing imagery datasets demonstrate the effectiveness of O-DINO, achieving state-of-the-art performance on DOTA, DIOR-R, and HRSC benchmarks. Our code will be released at https://github.com/liufeinuaa/rsisdet.git.