OSPIR-DETR: Object-Selective Proxy Interaction and Gain-Guided Scale Alignment for Remote Sensing Object Detection

Abstract

Remote sensing imagery is characterized by complex backgrounds, substantial scale variation, and numerous small targets, placing greater demands on semantic modeling of high-spatial-resolution features and feature integration across multiple scales. Within the D-FINE hybrid encoder, however, the relatively high-spatial-resolution feature receives limited global semantic interaction, while directly applying dense self-attention would incur substantial computational cost and would apply semantic updates uniformly across all spatial locations. Meanwhile, fixed cross-scale resampling does not explicitly model content-dependent local correspondence between adjacent feature levels. To address these issues, we propose OSPIR-DETR. First, we introduce Object-Selective Proxy Interaction Refinement (OSPIR), which aggregates global context from all spatial features through learnable proxy tokens while restricting semantic readout and residual updates to selected foreground candidate positions, thereby enabling global semantic modeling with selective feature refinement. Second, we propose Adjacent-Scale Gain-Guided Alignment Refinement (ASGAR), which estimates continuous sampling offsets from local correlation responses and modulates the offsets using center-referenced matching gains to improve local feature alignment before cross-scale fusion. Evaluation on the three benchmarks shows average precision (AP) gains of 1.1 points on DOTA-v1.0, 1.2 points on NWPU VHR-10, and 2.5 points on VEDAI over the matched D-FINE-L baseline, indicating consistent performance gains across the three benchmarks.