PS-TOD: Point-Supervised Tiny Object Detection via Two-Stage Calibration

Le résumé fourni par la source

Although fully supervised tiny object detection has achieved significant success, it relies on a large amount of manually annotated data as prior knowledge. To reduce annotation costs, point-supervised object detection (PSOD) has gained widespread attention. However, existing PSOD algorithms that directly generate pseudo boxes from point annotations still suffer from issues such as inaccurate bounding box regression, especially when dealing with tiny objects. To address this issue, this paper explores a novel point-supervised tiny object detection (PS-TOD) paradigm. Specifically, in the first stage, points are used to generate initial rough boxes; in the second stage, these rough boxes are treated as ground truth and further refined using robust object detection techniques to learn higher-quality detection boxes. We conduct experiments on two widely used tiny object detection datasets, AI-TOD and VisDrone, demonstrating promising results.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.