Rethinking the anchoring mechanism in oriented fine-grained object detection in remote sensing images: From rigid prior to discrimination-oriented anchor assignment

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

The spatial mismatch between pre-defined rigid anchors and the spatially uncertain discriminative regions of fine-grained objects severely hinders representation learning in remote sensing images. To address this, we propose a Geometry-Constrained Dynamic Anchor (GCDA) scheme to actively align candidates with discriminative features. Specifically, a Position Offset Calibration Module based on deformable convolutions is designed to adaptively shift anchor centers, leveraging a Gaussian mask to filter background noise. Furthermore, a Shape Evolution Adaptation Module is introduced to enforce geometric similarity between anchors and ground truths (i.e., side ratios and diagonal angles), enabling robust self-calibration of anchor shapes and orientations. Extensive experiments on VEDAI, HRSC2016, and ShipRSImageNet datasets demonstrate the effectiveness of GCDA, which attains mAPs of 77.8%, 74.3%, and 75.26%, respectively, demonstrating consistent performance across datasets with different scales and fine-grained category diversity.