Beyond Horizontal Paradigm: Circle Anchors for Oriented Object Detection

State-of-the-art oriented object detectors are predominantly built upon horizontal object detection paradigm, relying on horizontal anchors to generate oriented proposals or bounding boxes. However, these traditional horizontal anchor-based methods encounter the regression ambiguity problem, where different regression outcomes can produce identical oriented bounding boxes, hindering the training efficiency of the regression branch. To address this challenge, we propose a novel circle anchor-based representation for oriented object detection, which inherently circumvents regression ambiguity. Building on this foundation, we introduce circle anchor net, an effective two-stage oriented object detector. Furthermore, we develop the circle-to-square dynamic assigner, a straightforward yet effective strategy to compute overlaps between circle anchors and ground-truth boxes while dynamically adjusting sampling thresholds. Without bells and whistles, our method achieves state-of-the-art results on the recently released small oriented object detection dataset SODA-A [40.8% average precision (AP)], as well as on widely used benchmarks DOTA v2.0 (61.7% AP50), DIOR-R (73.4% AP50), and HRSC2016 (90.66% AP50), demonstrating its robustness and consistent superiority.