DFVNet: Density Feature Guided Small Object Detection in Video

Detecting small objects is a critical task in various application domains, including autonomous driving and drone recognition. However, existing small object detection methods, which are predominantly based on single images, often fail to capture essential information due to issues such as density and motion blur. These methods also neglect the temporal context inherent in real-world scenarios, resulting in suboptimal detection performance. To address these limitations, this paper proposes a novel video-based Detection Transformer Network, named DFVNet, for small object detection. DFVNet leverages density features to extract information from densely packed s and enhances the current detection frame through a coarse-to-fine refinement process, utilizing reference features from adjacent temporal frames. Extensive experiments on the VisDrone2019-VID dataset show that DFVNet improves average precision AP₅₀ by 3.0% compared to image-based detection methods and by 4.1% over other video-based detection methods.