Shape-Adaptive Learning and Style-Aware Decomposition for Domain-Generalized Oriented Object Detection

Oriented object detection is essential for interpreting high-resolution remote-sensing images, but detectors trained on a single source domain often suffer from severe performance degradation under unseen imaging conditions. Existing domain generalization methods mainly address horizontal detection or treat style variation as an undifferentiated perturbation, while the geometry-sensitive nature of oriented objects, especially elongated instances, is insufficiently considered. To address these limitations, a shape-adaptive learning and style-aware decomposition framework, termed SA-DGOD, is proposed for single-source domain-generalized oriented object detection. A frozen CLIP image encoder is employed only during training to provide multi-level style statistics, from which several interpretable branch-style pseudo-domains are constructed by modulating feature statistics that characterize photometric and contrast variations. These explicit pseudo-domain identities enable structured style diversification without accessing target-domain data. To improve orientation-sensitive representation learning, a shape-adaptive soft consistency loss is introduced for rotated RoI features. Instead of using a fixed-threshold binary weighting scheme, the proposed loss adopts an aspect-ratio-aware threshold and a continuous rotated-IoU-based weight, allowing elongated objects and geometrically reliable proposals to be handled more flexibly. Furthermore, a style-aware common-specific classifier decomposition is introduced into the proposed framework to separate domain-invariant category discrimination from pseudo-domain-specific style bias. Extensive experiments on multiple remote sensing benchmarks demonstrate the effectiveness and generalization capability of SA-DGOD.