A2Net: An Anchor-Free Alignment Network for Oriented Object Detection in Remote Sensing Images

Abstract

Object detection in remote sensing images is crucial for identifying and locating objects in the field, holding significance in remote sensing. Oriented object detection employs oriented bounding boxes to locate objects with varying orientations, achieving recent advancements. However, challenges persist due to vast variations in object scale and orientation. Existing methods use Intersection over Union (IoU) to measure bounding box quality but often ignore shape information. Unlike horizontal detectors, oriented detectors always incorporate an angle parameter. Yet, objects with different shapes exhibit varying angle sensitivity. For objects with the same angle but different shapes, their IoU can differ significantly. We argue that relying solely on IoU is not comprehensive. To address this, we propose the Shape-aware IoU Score (SaIS), considering shape information and IoU for each bounding box. We use SaIS to enhance the dynamic soft label assignment strategy, resulting in an improved Shape-aware Label Assignment (SaLA). SaLA aids the detector in selecting more suitable samples. Leveraging RTMDet-R and S2ANet strengths, we design an Anchor-free Alignment Network (A2Net) for oriented object detection. A2Net features two detection heads: the initial head and the refinement head. Utilizing alignment convolution (AlignConv) between these heads obtains aligned features. We validate the proposed approach’s effectiveness on the DOTA dataset and DIOR-R dataset.

Keywords

• Oriented object detection
• shape-aware
• remote sensing
• deep learning