Position-DETR: Step-by-Step Position-Guided Small Object Detection in Remote Sensing Images

Detecting small objects in remote sensing images is highly challenging, primarily due to background noise interference and the limited semantic information of small objects. Although detection transformer (DETR)-like methods have significantly improved detection performance by introducing the transformer architecture, their performance remains suboptimal for small target detection. This is primarily because these methods struggle to effectively focus on small objects, leading to an equal treatment of extensive background information and sparse foreground information. This phenomenon not only results in the oversight of small objects but also introduces unnecessary computational burdens and slower convergence speeds. To address the above problems, we propose Position-DETR. First, we design a hierarchical foreground selection (HFS) module and a fine-grained foreground enhancement (FGFE) module to resolve computational bias and redundancy in the encoder. Furthermore, we propose progressive attention optimization (PAO) and hybrid matching query (HMQ) in the decoder to supplement small object information and improve convergence speed. Experimental results on four public remote sensing datasets for small target detection (DOTA-v1.0, DIOR, VISDRONE-2019, and STAR) demonstrate the effectiveness of Position-DETR, achieving average precision (AP) scores of 42.3%, 65.5%, 25.3%, and 21.3%, respectively, surpassing previous state-of-the-art results on each dataset. Our method achieves significantly faster convergence while maintaining low computational complexity. The source code will be available at https://github.com/wenkaizhao/Position-DETR