Recursive Gate-Controlled Convolution for Remote Sensing Detection

Remote sensing (RS) multi-class object detection holds great potential in various fields such as agriculture production, urban planning and military reconnaissance. Recent research on RS object detection has primarily focused on utilizing multi-scale feature while overlooking the unique prior knowledge in global information. This prior knowledge can be valuable since small RS objects may be incorrectly detected without sufficient reference to global information, and different types of objects may require different global information. In this paper, we address this prior factor by introducing recursive gate-controlled convolution (RGC). RGC possesses the capability to model global information, allowing for the recursive detection of spatially adjacent features related to the objects through simple operations. By doing so, it better emulates the detection performance of various objects in RS scenario. Additionally, recognizing the crucial role played by multi-scale features in dense scenes, we have preserved this knowledge and enhanced the interaction between features using the bidirectional feature pyramid. Our experimental results on the FAIR1M have achieved competitive performance (68.0% mAP), significantly surpassing the methods employed in FAIR1M.