Align Deep Features for Oriented Object Detection

The past decade has witnessed significant progress on detecting objects in aerial images that are often distributed with large-scale variations and arbitrary orientations. However, most of existing methods rely heuristically defined anchors different scales, angles, aspect ratios, usually suffer from severe misalignment between anchor boxes (ABs) axis-aligned convolutional features, which lead to the common inconsistency classification score localization accuracy. To address this issue, we propose a single-shot alignment network (S 2 A-Net) consisting two modules: feature module (FAM) an oriented detection (ODM). FAM can generate high-quality refinement network adaptively align features according ABs novel convolution. ODM first adopts active rotating filters encode orientation information then produces orientation-sensitive orientation-invariant alleviate Besides, further explore approach detect large-size images, leads better trade-off speed Extensive experiments demonstrate our method achieve state-of-the-art performance commonly used objects’ data sets (i.e., DOTA HRSC2016) while keeping high efficiency.