Feature Alignment FPN for Oriented Object Detection in Remote Sensing Images

Abstract

As the basic and essential component of most object detectors, a feature pyramid network (FPN) can effectively extract multiscale features to recognize objects at different scales. Nevertheless, due to the cross-scale fusion and upsampling operations in FPN, current detectors still suffer from information loss and feature misalignment. To alleviate these problems, a flow-guided upsampling module (FGUM) and a multifeature attention module (MFAM) are proposed to improve the FPN. Specifically, FGUM uses a novel flow warp in the upsampling operation to align features, resulting in better cross-scale fusion. At the same time, the MFAM module fully considers the integrity of multiscale features and reduces the aliasing effect by optimizing the weight of each level of features. To verify the effectiveness of the improved FPN, it is applied to three state-of-the-art models for experiments, and all of them achieve better detection accuracy compared to the original FPN.