Rotated Feature Network for Multiorientation Object Detection of Remote-Sensing Images

Abstract

Modern detectors in remote-sensing images follow the pipeline that feature maps extracted from ConvNets are shared between classification and regression tasks. However, there exist obvious conflicting demands in multiorientation object detection of remote-sensing images (RSOD) that classification is insensitive to orientations, while regression is quite sensitive. In addition, previous works cannot promise the reliability of rotation-invariant or rotation-equivariant features with only qualitative or intuitive analysis. To address these issues, we propose an encoder-encoder architecture, called rotated feature network (RFN), which produces rotation-sensitive feature maps (RS) for regression and rotation-invariant feature maps (RI) for classification. Specifically, the encoder unit assigns weights for rotated feature maps. The decoder unit extracts RS and RI by performing resuming operators on rotated and reweighed feature maps, respectively. To make the rotation-invariant characteristics more reliable, a metric is adopted to quantitatively evaluate the rotation-invariance by adding a constraint item in the loss, yielding a promising detection performance. Compared with the state-of-the-art methods, our method can achieve a significant improvement in NWPU very high resolution (VHR)-10 and RSOD data sets. The proposed RFN is further evaluated on the scene classification in remote-sensing images, demonstrating its good generalization ability. It can be integrated into an existing framework, leading to better performance with only a slight increase in test time and model complexity.