AgriShelf: A Multi-Class, Bi-Source Image Dataset for Smart Agri-Food Retailing Applications

In this dataset, we have compiled a comprehensive collection of 16,592 agri-food retail images across various classes commonly found in grocery and supermarket environments. To ensure generalizability, the dataset was collected using two distinct sources: a smartphone and an Intel RealSense Depth Camera (D435i), under diverse, real-world conditions, such as shelf inclinations, lighting levels, and different angles. The dataset is structured into two main subsets: unlabeled and labeled. The unlabeled subset is curated for key computer vision tasks relevant to retail applications, including classification, object detection, and product recognition. The labeled subset consists of 2,416 samples with detailed centroid annotations, making it suitable for On-Shelf Availability (OSA) estimation, counting, or multi-task learning approaches. Altogether, both subsets serve as valuable benchmarks for evaluating and testing automated inventory monitoring systems and real-time retail analytics applications.本数据集收录了16592幅农业食品零售图像，涵盖商超与杂货店场景中常见的各类商品品类，形成了一套全面的图像集合。为保障数据集的泛化性能，本数据集通过两种差异化采集渠道构建：智能手机与英特尔实感深度相机（Intel RealSense Depth Camera D435i），采集场景覆盖多样的真实环境变量，包括货架倾斜状态、光照条件与拍摄视角等。 本数据集分为两大核心子集：未标注子集与标注子集。 未标注子集经过精心筛选，适配零售场景相关的关键计算机视觉任务，涵盖图像分类、目标检测与商品识别等方向。 标注子集包含2416个样本，附带详细的中心点标注，可适用于货架商品可用性（On-Shelf Availability，OSA）评估、商品计数或多任务学习等研究场景。 综上，两大子集均可作为评估与测试自动化库存监控系统及实时零售分析应用的优质基准数据集。