SA3Det: Detecting Rotated Objects via Pixel-Level Attention and Adaptive Labels Assignment

Abstract

Remote sensing of rotated objects often encounters numerous small and dense objects. To tackle small-object neglect and inaccurate angle predictions in elongated objects, we propose SA3Det, a novel method employing Pixel-Level Attention and Adaptive Labels Assignment. First, we introduce a self-attention module that learns dense pixel-level relations between features extracted by the backbone and neck, effectively preserving and exploring the spatial relationships of potential small objects. We then introduce an adaptive label assignment strategy that refines proposals by assigning labels based on loss, enhancing sample selection during training. Additionally, we designed an angle-sensitive module that enhances angle prediction by learning rotational feature maps and incorporating multi-angle features. These modules significantly enhance detection accuracy and yield high-quality region proposals. Our approach was validated by experiments on the DOTA and HRSC2016 datasets, demonstrating that SA3Det achieves mAPs of 76.31% and 89.4%, respectively.