SPMR: Bridging Set Prediction and Multipositive Recognition for Fine-Grained Object Detection in Remote Sensing Images

Fine-grained object detection (FGOD) in remote sensing requires accurate localization and fine-grained recognition under subtle visual differences. Current query-based detectors enable NMS-free inference through one-to-one set prediction, but their sparse positive supervision makes fine-grained recognition difficult. To mitigate this limitation, we propose SPMR, a framework that combines set prediction with multi-positive recognition for remote sensing FGOD. SPMR separates class-agnostic proposal generation from fine-grained recognition and applies different supervision to the two stages. The Sparse Oriented Proposal Network (SOPN) generates sparse class-agnostic oriented proposals and estimates their objectness under one-to-one set supervision. The Query-Aware Refinement Head (QARH) then performs proposal-level fine-grained recognition with quality-aware multi-positive supervision. During inference, Dual-Stage Consensus Scoring (DSCS) fuses the objectness score from SOPN with the fine-grained class score from QARH to produce final predictions without NMS. Experiments on MAR20, ShipRSImageNet, and FAIR1M show that SPMR outperforms representative NMS-free oriented detectors and remains competitive with strong two-stage FGOD methods. Specifically, SPMR obtains 85.82 AP on MAR20, 75.40 AP on ShipRSImageNet, and 52.12 AP on FAIR1M-v2.0. Code is available at: https://github.com/AIRVIC-Lab/SPMR.