ELFT: Efficient local-global fusion transformer for small object detection

Small object detection is an essential but challenging task in computer vision. Transformer-based algorithms have demonstrated remarkable performance in the domain of computer vision tasks. Nevertheless, they suffer from inadequate feature extraction for small objects. Additionally, they face difficulties in deployment on resource-constrained platforms due to their heavy computational burden. To tackle these problems, an efficient local-global fusion Transformer (ELFT) is proposed for small object detection, which is based on attention and grouping strategy. Specifically, we first design an efficient local-global fusion attention (ELGFA) mechanism to extract sufficient location features and integrate detailed information from feature maps, thereby promoting the accuracy. Besides, we present a grouped feature update module (GFUM) to reduce computational complexity by alternately updating high-level and low-level features within each group. Furthermore, the broadcast context module (CB) is introduced to obtain richer context information. It further enhances the ability to detect small objects. Extensive experiments conducted on three benchmarks, i.e. Remote Sensing Object Detection (RSOD), NWPU VHR-10 and PASCAL VOC2007, achieving 95.8%, 94.3% and 85.2% in mean average precision (mAP), respectively. Compared to DINO, the number of parameters is reduced by 10.4%, and the floating point operations (FLOPs) are reduced by 22.7%. The experimental results demonstrate the efficacy of ELFT in small object detection tasks, while maintaining an attractive level of computational complexity.