CoffeeBeanSeg: A Polygon-Annotated Instance Segmentation and Quality Grading Dataset with YOLOv11 Benchmarking

This work focuses on the clear and efficient assessment of pre-roast coffee bean quality in the global agricultural supply chain. This task faces significant challenges. While computer vision is a scalable method, progress has been limited owing to the general rarity of specialist public datasets. Most current collections tend to focus more on coarse classification rather than the fine-grained localization necessary for automated sorting. To fill this gap, we present a new, publicly available dataset of 3,877 high-resolution images of raw beans of Coorg origin. Manual classification of these beans was done by local experts into four quality grades: A- Premium, B- Good, C-Standard, and D- Defective. The main contribution of the paper is the subset of 2,284 images with careful annotation of their exact polygon masks and grade labels, thereby allowing advanced instance segmentation model development. Data acquisition was done under controlled lighting with multiple consumer-grade devices to ensure practical relevance of results. We outline a comprehensive protocol for data collection and verify the quality of the dataset by establishing a strong baseline using a state-of-the-art YOLOv11 model. This dataset is designed to support research in automated agricultural grading and to help create models that can perform object detection, segmentation, and classification at the same time.