Deep Spatial Feature Transformation for Oriented Aerial Object Detection

Object detection in aerial images has received extensive attention in the field of computer vision. Different from natural images, the aerial objects are usually distributed in any direction. Therefore, the existing detector usually needs more parameters to encode the direction information, resulting in a large number of redundant calculations. In addition, because an ordinary convolution neural network (CNN) does not effectively model the direction change, a large amount of the rotated data is required for the aerial detector. To solve these problems, we propose a deep spatial feature transformation network (DSFT-Net), which includes a spatial feature extraction module and a feature selection module. Specifically, we add the rotation convolution kernel to the detector to extract the directional feature of the rotated target to accurately predict the direction of the model. Then, we build a dual pyramid to separate the features in the classification and regression tasks. Finally, the polarization function is proposed to construct the critical features that are suitable for their respective tasks, achieving feature selection and more refined detection. Experiments on public remote sensing benchmarks (e.g., DOTA, HRSC2016, and UCAS-AOD) have proved the effectiveness of our detector.