EVENT DOSSIER
On September 7, 2026, researchers proposed a new setting for Group-Individual Object Counting (GIC) on arXiv, requiring the model to simultaneously count individual objects and semantic groups within the same framework. To this end, the team released the BunchCount benchmark dataset containing 1,330 images, 89,254 individual annotations, and 1,330 group annotations, and clarified the inclusion relationships between groups and individuals. Experiments showed that existing advanced models performed well in counting individual instances, but lacked accuracy in counting semantic groups. To address this issue, the research team proposed a relational counting framework guided by counting units, using the inclusion relationships between groups and individuals to regularize representations at different granularities. This method significantly improved group-level counting capabilities while maintaining individual-level counting performance, establishing a strong benchmark for counting tasks beyond individual instances.
研究人员提出 Group-Individual Object Counting (GIC) 新设定，要求模型在统一框架下同时计数个体对象与语义群组。为此，团队发布 BunchCount 基准数据集，包含 1,330 张图像、89,254 个个体标注及 1,330 个群组标注，并明确记录群组与其构成个体的包含关系。实验表明，现有先进计数模型在个体实例上表现良好，但在语义群组计数上准确性不足。为此，研究团队提出计数单元引导的关系计数框架，利用群组与个体的包含关系正则化跨粒度表示。该方法显著提升了群组级计数能力，同时更好地保留了个体级计数性能，为超越实例的计数任务建立了强基准。