One-Stage Cascade Refinement Networks for Infrared Small Target Detection

Single-frame InfraRed Small Target (SIRST) detection has been a challenging task due to lack of inherent characteristics, imprecise bounding box regression, scarcity real-world datasets, and sensitive localization evaluation. In this paper, we propose comprehensive solution these challenges. First, find that the existing anchor-free label assignment method is prone mislabeling small targets as background, leading their omission by detectors. To overcome issue, an all-scale pseudo-box-based scheme relaxes constraints on scale decouples spatial from size ground-truth target. Second, motivated structured prior feature pyramids, introduce one-stage cascade refinement network (OSCAR), which uses high-level head soft proposals for low-level head. This allows OSCAR process same target in coarse-to-fine manner. Finally, present new research benchmark infrared detection, consisting SIRST-V2 dataset real-world, high-resolution single-frame targets, normalized contrast evaluation metric, DeepInfrared toolkit detection. We conduct extensive ablation studies evaluate components compare its performance state-of-the-art model-driven data-driven methods benchmark. Our results demonstrate top-down framework can improve accuracy without sacrificing efficiency. The toolkit, dataset, trained models are available at https://github.com/YimianDai/open-deepinfrared.