DARDet: A Dense Anchor-Free Rotated Object Detector in Aerial Images

Rotated object detection in aerial images has received increasing attention for a wide range of applications. However, it is also challenging task due to the huge variations scale, rotation, aspect ratio, and densely arranged targets. Most existing methods heavily rely on large number predefined anchors with different scales, angles, ratios, are optimized distance loss. Therefore, these sensitive anchor hyperparameters easily suffer from performance degradation caused by boundary discontinuity. To handle this problem, letter, we propose dense anchor-free rotated detector (DARDet) images. Our DARDet directly predicts five parameters boxes at each foreground pixel feature maps. We design new alignment convolution module (ACM) extract aligned features introduce pixels-intersection over union (PIoU) loss precise stable regression. method achieves state-of-the-art three commonly used objects datasets (i.e., DOTA, HRSC2016, UCAS-AOD) while keeping high efficiency. Code available https://github.com/zf020114/DARDet .