Rotation-Aware Adaptor via Equivariance-Constrained Consistency Regularization for Remote Sensing Detection

Abstract

Remote sensing has become indispensable in applications such as urban planning, maritime surveillance, and disaster management, yet accurate object detection remains challenging due to arbitrary orientations, dense layouts, and large intraclass variations. Existing oriented detectors have made progress through rotation-sensitive heads and feature-alignment mechanisms, but they still handle rotation primarily in an implicit manner, lacking explicit geometric priors and often exhibiting inconsistency between feature representations and final predictions. To address these limitations, we design a parameter-efficient modular adaptor readily applicable to diverse detectors with only a small learnable-parameter overhead. Specifically, we introduce a rotation-aware multi-mother convolution (RAMMC), which explicitly embeds rotation as a structural inductive bias: a multibranch adaptor enhances rotation-aware feature construction, while two complementary regularization terms enforce feature-level equivariance and prediction-level consistency, ensuring coherent and geometry-aligned detection behavior. The design is inspired by rotation equivariance and invariance theory. We provide a conditional theoretical analysis showing that the group-pooling branch of RAMMC realizes a Reynolds averaging operator over the discrete rotation group, and we derive an implementation-aware equivariance-defect bound that accounts for finite kernel rotation, spatial gating, and orientation-attention residuals. Extensive experiments on diverse detectors, such as Faster R-CNN, RoI Transformer, and Oriented R-CNN implemented in the MMRotate toolbox, as well as YOLO models implemented under the Ultralytics framework, confirm that our adaptor consistently improves performance. Notably, it achieves 84.32 mean Average Precision (mAP) on DOTA-v1.0, 47.31 mAP on FAIR1M-v1, and 52.28 mAP on FAIR1M-v2.0, surpassing strong baselines. These results demonstrate that explicitly modeling rotation consistency through a parameter-efficient adaptor with mathematically grounded rotation priors and consistency constraints provides an effective and general solution for remote sensing object detection.