Pas: a scale-invariant approach to maritime search and rescue object detection using preprocessing and attention scaling

Abstract

Object detection is a primary means of unmanned aerial vehicle (UAV) maritime search and rescue. The problem of scale variation caused by UAV flight height changes, shooting angle changes, and giant waves seriously affects the detection performance. However, most work does not explicitly consider the effects of these factors. In this work, we propose an algorithm called Preprocessing and Attention Scaling, which explicitly considers the scale variation problem caused by height, angle changes, and giant waves for the first time and solves it through Preprocessing Scaling and Attention Scaling. The Preprocessing Scaling module scales and perspective changes the images according to each photograph’s recorded flight altitude and shooting angle and crops them to the appropriate size, significantly improving the detection accuracy and shortening the inference time. At the same time, the scale variation caused by the up and down of the object due to the vast swells cannot be solved by the Preprocessing Scaling module, so we designed the Attention Scaling module again to quickly capture the area that needs further scale change by fusing the horizontal attention and vertical attention, and then transform it to the appropriate scale by the affine transformation, further improving detection accuracy. We extensively tested PAS on the well-known SeaDronesSee-DET and the SeaDronesSee-DET v2 (S-ODv2) datasets, significantly improving the detection accuracy. In addition, we successfully tested our method on a height-angle transfer task, where we trained on some height-angle intervals and tested on different height-angle intervals, achieving good results.