Harnessing Attention Mechanisms for Improving Small Object Detection in Drone-Captured Scenarios

At present, object detection in drone-captured scenarios has gained significant popularity due to its wide-ranging applications. However, existing object detection methods face challenges, particularly detecting small objects due to the scale variations, dense arrangements, and sparse distribution of objects. This paper addresses the issue by focusing on improving the backbone network of baseline detection model and aims to improve the feature extraction capability of the model. To achieve this, we conducted this study in two key steps: firstly, replacing conventional convolution blocks with attention modules (SE, ECA, CA and CBAM) in backbone of baseline model network. Secondly, evaluation of the effectiveness of added attention modules on VisDrone dataset, followed by the comparative analysis of results with baseline model. The results demonstrate the efficacy of attention modules by increasing detection accuracy, simultaneously reducing computational complexity (FLOPs) and number of parameters.