Adaptive Dynamic Anchors for Small Object Detection in Remote Sensing Imagery

Small object detection in remote sensing imagery remains a challenging task due to the dense spatial distribution of objects and the scarcity of positive samples during training. While existing techniques such as data augmentation and adaptive sample assignment strategies offer partial relief, they do not fundamentally address the core issue of anchor-object mismatch. In this work, we present a novel detection framework that leverages adaptive dynamic anchors to improve detection precision for small objects. Our method consists of three components: an adaptive anchor shape generation module that automatically generates anchors whose shapes better match the actual distribution of object sizes in the dataset, a feature-guided dynamic anchor localization module that allows more precise spatial alignment between anchors and objects, and a size-aware dynamic sample assignment strategy that ensures sufficient positive samples for small-scale objects. The experimental results highlight the effectiveness of our approach in bridging the gap between anchor design and object characteristics in remote sensing imagery. The proposed Adaptive Dynamic Anchor Network (ADANet) achieves $38.2 \% \mathrm{mAP}$ on the SODA-A dataset, demonstrating notable improvements over state-of-the-art methods.