Improved Transformer-Based SSD Detector for Airborne Object Detection

Object detection tasks based on UAV -captured images are often made quite challenging because their images contain great scale variance, a high number of interferers and complex backgrounds. In our work, a modified CSP-DarkNet backbone is designed to extract features with greater variance while ensuring sufficient global information. We introduce an improved BiFPN for more effective multi-scale feature fusion. Our experimental results show that the SSD-VIT significantly outperforms state-of-the-art detectors and achieves excellent results in the VisDrone-DET2021 challenge.