3DF-FCOS: Small object detection with 3D features based on FCOS

Abstract

Lower resolution and fewer feature details appear in small objects, so achieving small object detection is a challenging task. To improve the detection ability in the case of low-resolution object details, we propose to exploit 3D features with FCOS for representing small objects. The proposed method employs a Global Spatial Block (GS-Block) in the backbone to guide the network to learn the picture’s shallow features and increase the model’s perceived capability for small objects. In addition, we introduce a 3D sparse convolution in the shallow input features of the detection heads to better capture the global and contextual information of the shallow feature elements and further improve the feature representation of small objects in the backgrounds, and the whole network is referred to as 3DF-FCOS. Finally, we conducted extensive experiments to verify the effectiveness of 3DF-FCOS on VisDrone2019Det. The results show that our method achieves a remarkable improvement in AP and APs compared to the baselines of 1.2% and 1.2%, respectively. Compared with current state-of-the-art (SOTA) algorithms, the performance of our proposed 3DF-FCOS is comparable.