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

Description

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.

Steps to reproduce

For the smartphone, this dataset was collected using an iPhone 14 Plus at 1080p resolution and 30 frames per second with High Dynamic Range (HDR) across different agri-food retailers. We recorded multiple videos (approximately 20 to 30 seconds each). Then, we converted them from MP4 to JPG at 10 frames per second. As for the Intel RealSense Depth Camera (435i), an automatic timer was set to collect real-time images from the camera every 10 seconds.

Institutions

Categories

Funding

Licence