Effective Anchor Adaptation and Feature Enhancement Strategies for Tiny Object Detection in Aerial Images

Abstract

In recent years, research based on anchor-based two-stage detectors has achieved great performance improvements in aerial object detection tasks. However, they still have two significant problems in the detection of tiny objects: 1) The preset fixed anchor is not conducive to assigning positive and negative samples in RPN when dealing with tiny objects, resulting in low-quality samples. 2) When the detector encounters tiny objects lacking structural details, it fails to accurately represent features, causing divergence in object features and hindering network learning. In this work, we propose the Anchor Adaptation and Feature Enhancement Strategies (AFS) to alleviate the above two problems. AFS contains two optimized modules: Anchor Adaption RPN Head (A2RH) and Feature Enhanced Attention Module (FEAM). Specifically, A2RH performs anchor adaptive learning by establishing a new anchor bias learning branch from the feature map, enabling higher-quality positive and negative sample assignments in RPN. FEAM introduces global features and mask attention based on FPN, and presents Gaussian mask supervision for attention to obtain stronger feature representation. Experiments show that our method improves the average precision by 1.8% on the baseline model, and achieves state-of-the-art results on AI-TOD dataset. Moreover, validation on AI-TOD-v2 and VisDrone2019 datasets also confirms the effectiveness of our method. The code will soon be available at https://github.com/gravity-lhg/AFS.

Keywords

• Deep learning
• aerial images
• tiny object detection
• anchor adaptation
• feature enhancement