RAPTOR: Rotational Adaptive Parallel Topology for Object Detection in remote sensing

Military Remote Sensing Object Detection (MRSOD) is inherently challenged by arbitrarily oriented objects, adversarial camouflage, and atmospheric interference. When tackling these issues, conventional CNNs struggle with rotational feature misalignment, while Transformers incur prohibitive quadratic computational costs. To overcome these limitations, we propose RAPTOR (Rotational Adaptive Parallel Topology for Object Detection), a highly efficient backbone architecture. At its core, the Parallel Local–Global Fusion Block (PLGFB) synergizes a linear-complexity Vision Mamba for holistic context modeling with a Group-wise Rotational Deformable Convolution (GR-DC) for adaptive local geometric alignment. To alleviate spatial information loss for micro-instances, we design an Adaptive Dynamic-Routed Fusion Down-sampling (ADRFD) module. Furthermore, a Laplacian-of-Gaussian Guided Stem (LoGGS) is introduced to suppress noise and enhance edge priors. To address the scarcity of fine-grained benchmarks, we constructed Military-RSOD, a high-quality dataset comprising 53 distinct military categories. Extensive experiments on Military-RSOD and DOTA-v1.0 demonstrate that RAPTOR achieves state-of-the-art performance. Notably, it delivers an impressive 86.39% mAP with a low computational footprint of only 53.23 G FLOPs, presenting a highly practical equilibrium between accuracy and parameter efficiency. The source code and dataset are available at https://github.com/KeLiu-y/RAPTOR.