Multi-factor consideration sample assignment for oriented tiny object detection

Tiny object detection in aerial image is crucial for urban planning and environmental monitoring. However, unpredictable orientation and lack of distinctive features pose challenges in sample assignment, often resulting in mismatch and inconsistency between anchors and priors. To address this, we introduce the multi-factor consideration sample assignment (MCSA) mechanism, which ensures the assignment of superior positive samples for objects with orientation. Initially, we craft a dynamic prior block (DPB) to facilitate the dynamic alignment of priors with objects. Subsequently, we introduce an anchor assessment metric that assesses all potential anchors thoroughly. Lastly, we deploy a dynamic Gaussian mixture model (DGMM) to eliminate subpar samples. Our method outperforms most current techniques on four datasets: DIOR-R, DOTA-v2.0, DOTA-v1.5 and DOTA-1.0. Notably, we achieve the mAP of 51.86% on the DOTA-v2.0 dataset, surpassing the baseline by 5.18 percentage points. By distributing priors dynamically and selecting the most compatible positive samples based on the highest matching scores, our approach ensures precise sample assignment, consequently enhancing detection precision.