Remote sensing object detection remains a formidable challenge in computer vision, particularly for arbitrarily oriented objects in complex scenes. Representative points-based methods, such as RepPoints, have emerged as a robust alternative to angle regression, yet they suffer from mutual interference between classification and localization tasks due to shared point sets. To address this underexplored issue, we propose Center Corrective Representative Points (CCRP), a novel framework that decouples these tasks through three innovative components: (1) a dynamic center correction strategy that computes an adaptive center point from representative points’ mean to extract robust classification features, (2) a boundary-aware loss that penalizes deviations from object boundaries for precise localization, and (3) an adaptive sample selection mechanism with smoothed weighting to prioritize high-quality samples during training. Unlike anchor-based methods requiring extensive tuning or recent Gaussian-based approaches with high computational costs, CCRP offers a lightweight, cohesive solution. Extensive experiments on four challenging datasets—DOTA, HRSC2016, UCAS-AOD, and DIOR-R—demonstrate CCRP’s superiority, achieving mAP improvements of 2.14%, 0.94%, 1.82%, and 1.55% over baselines, respectively, underscoring its innovation and efficiency.