OPT-Net: An Orientation-Preserving Transformer for End-to-End Oriented Object Detection in Remote Sensing Images

The objects in high-resolution remote sensing images usually exhibit arbitrary orientations, multi-scale variations, dense distributions, and complex background interference, posing significant challenges to oriented object detection. Although existing DETR-style end-to-end detectors eliminate the need for anchor design and non-maximum suppression, they still suffer from insufficient orientation priors in object queries, limited orientation consistency in decoder feature interaction, and unstable set matching for oriented bounding boxes. To address these issues, this paper proposes an end-to-end Transformer framework, termed OPT-Net (Orientation-Preserving Transformer Network), for oriented object detection in remote sensing images. OPT-Net treats orientation information as a structured geometric prior and propagates it through query initialization, feature interaction, and matching optimization. Specifically, an Orientation-Aware Query Initialization (OAQI) module is designed to generate initial queries using center confidence and orientation priors. An Orientation-Consistent Cross-Attention (OCCA) mechanism is proposed to perform orientation-conditioned modulation on Value features while keeping the standard Query–Key attention computation unchanged. Furthermore, an Uncertainty-aware Matching Loss (UML) is introduced to incorporate instance-level geometric uncertainty into Hungarian matching and regression optimization. Experimental results on the DOTA-v1.0 and HRSC2016 datasets show that OPT-Net achieves 76.83% and 90.58% mAP, respectively, demonstrating competitive detection accuracy and adaptability to complex remote sensing scenarios. Ablation studies and visualization results further validate the effectiveness of each proposed module.