Global Context-Anchored Cyclic Detector With Angle-Aware Regression for Remote Sensing Object Detection

Abstract

Remote sensing object detection (RSOD) is crucial for disaster monitoring, emergency response, and military reconnaissance, yet it remains challenging due to extreme scale variations and arbitrary orientations of objects. Despite recent advancements, existing methods are hindered by two critical challenges. At the feature representation level, the local progressive interaction paradigm in current feature pyramids causes semantic attenuation during long-range cross-scale transmission. At the regression supervision level, while Gaussian-based methods successfully bypass angular periodicity issues, their isotropic nature causes angular supervision to degrade when bounding boxes approach square shapes. To address these challenges, we propose the global context-anchored cyclic detector with angle-aware regression (GCAD). In particular, we introduce a global context-anchored cyclic feature pyramid network (GCAC-FPN) that employs an evolving unified context embedding spanning all scales. Through a recurrent local-to-global (L2G) and global-to-local (G2L) bidirectional interaction mechanism, this context embedding and multiscale local features continuously coevolve, enabling direct cross-scale information flow and feature refinement. Furthermore, we design an angle-aware regression (AAR) strategy that introduces explicit angular supervision into the regression loss, thereby maintaining orientation sensitivity when the Gaussian representations of square objects degenerate into isotropic distributions. Extensive experiments on two public benchmarks demonstrate that GCAD achieves state-of-the-art performance, reaching 80.72% mean average precision (mAP) on DOTA-v1.0 and 71.84% on DIOR-R.