MEDISEG

<b>Dataset Overview</b>This repository contains two related datasets for pill recognition research: MEDISEG and MEDISEG-Deploy.MEDISEG (MEDication Image SEGmentation) is a high-quality, real-world dataset designed for the development and evaluation of pill recognition models. It contains two subsets:MEDISEG (3-Pills): A controlled dataset featuring three pill types with subtle differences in shape and color.MEDISEG (32-Pills): A more diverse dataset containing 32 distinct pill classes, reflecting real-world challenges such as occlusions, varied lighting conditions, and multiple medications in a single frame.MEDISEG-Deploy is a deployment-oriented few-shot evaluation subset designed to support realistic pill recognition under visual domain shift. It includes cluttered multi-pill scenes with overlapping pills, varied backgrounds, pill boxes, and reflective surfaces, and provides fixed support and query partitions for few-shot adaptation under 1-shot, 5-shot, and 10-shot settings.Each dataset includes COCO-format annotations with instance segmentation masks, bounding boxes, and class labels.<b>Dataset Structure</b>The dataset is organized as follows:<br><br>MEDISEG/│── LICENSE<br>│── metadata.csv│── 3pills/<br>│ ├── annotations.json│ ├── images/<br>│ │ ├── image1.jpg│ │ ├── image2.jpg<br>│── 32pills/│ ├── annotations.json<br>│ ├── images/│ │ ├── image1.jpg<br>│ │ ├── image2.jpg<br>MEDISEG-Deploy/<br>│── LICENSE<br>│── 1-shot/<br>│ ├── annotations.json│ ├── images/<br>│ │ ├── image1.jpg│ │ ├── image2.jpg<br>│── 5-shot/│ ├── annotations.json<br>│ ├── images/│ │ ├── image1.jpg<br>│ │ ├── image2.jpg<br>│── 10-shot/<br>│ ├── annotations.json│ ├── images/<br>│ │ ├── image1.jpg│ │ ├── image2.jpg<br>│── query/│ ├── annotations.json<br>│ ├── images/│ │ ├── image1.jpg<br>│ │ ├── image2.jpg<br>LICENSE: The CC BY 4.0 license under which the dataset is distributed.metadata.csv: Supplementary drug information, including registration numbers, brand names, active ingredients, regulatory classifications, and official URLs.annotations.json: COCO-format annotation files providing segmentation masks, bounding boxes, and class labels.images/: High-resolution JPG images of medications.<b>Acknowledgements</b>If you use this dataset, please cite the corresponding publication:For MEDISEG (3-Pills and 32-Pills): <br>bibtex<br>@article{MEDISEG2026,<br>title = {A dataset of medication images with instance segmentation masks for preventing adverse drug events},<br>author = {Chu, Wai Ip and Hirani, Shashi and Tarroni, Giacomo and Li, Ling},<br>journal={arXiv preprint arXiv:2603.10825},<br>year = {2026},<br>doi={10.48550/arXiv.2603.10825}<br>}For MEDISEG-Deploy: <br>bibtex<br>@article{MEDISEGFewShot2026,<br>title = {Evaluating Few-Shot Pill Recognition Under Visual Domain Shift},<br>author = {Chu, Wai Ip and Tarroni, Giacomo and Li, Ling},<br>journal={arXiv preprint arXiv:2603.10833},<br>year = {2026},<br>doi={10.48550/arXiv.2603.10833}<br>}For work using both the core MEDISEG dataset and MEDISEG-Deploy: please cite both papers.<br><b>数据集概览</b> 本仓库包含两个用于药片识别研究的相关数据集：MEDISEG与MEDISEG-Deploy。其中MEDISEG（全称：MEDication Image SEGmentation，即药物图像分割数据集）是高质量的真实世界数据集，旨在开发与评估药片识别模型。该数据集包含两个子集： 1. MEDISEG（3-Pills）：受控数据集，涵盖三种在形状与颜色上存在细微差异的药片类型。 2. MEDISEG（32-Pills）：更多样化的数据集，包含32个不同的药片类别，能够反映真实场景中的各类挑战，例如遮挡、多变光照条件以及单帧内存在多种药物的情况。 MEDISEG-Deploy是面向部署的少样本（few-shot）评估子集，用于支持视觉域偏移下的真实药片识别任务。该子集包含杂乱的多药片场景，存在药片重叠、多样背景、药盒以及反光表面，并针对1样本、5样本和10样本的少样本适配任务提供了固定的支持集与查询集划分。 两个数据集均采用COCO格式（COCO-format）的标注文件，包含实例分割掩码、边界框与类别标签。 <b>数据集结构</b> 本数据集的组织形式如下： MEDISEG/ ├── LICENSE ├── metadata.csv ├── 3pills/ │ ├── annotations.json │ ├── images/ │ │ ├── image1.jpg │ │ ├── image2.jpg ├── 32pills/ │ ├── annotations.json │ ├── images/ │ │ ├── image1.jpg │ │ ├── image2.jpg MEDISEG-Deploy/ ├── LICENSE ├── 1-shot/ │ ├── annotations.json │ ├── images/ │ │ ├── image1.jpg │ │ ├── image2.jpg ├── 5-shot/ │ ├── annotations.json │ ├── images/ │ │ ├── image1.jpg │ │ ├── image2.jpg ├── 10-shot/ │ ├── annotations.json │ ├── images/ │ │ ├── image1.jpg │ │ ├── image2.jpg ├── query/ │ ├── annotations.json │ ├── images/ │ │ ├── image1.jpg │ │ ├── image2.jpg 各文件说明如下： - LICENSE：本数据集采用的CC BY 4.0开源许可协议。 - metadata.csv：补充药物信息，包含注册编号、商品名称、活性成分、监管分类以及官方网址。 - annotations.json：COCO格式（COCO-format）的标注文件，提供实例分割掩码、边界框与类别标签。 - images/：高分辨率JPG（JPEG）格式的药物图像。 <b>致谢</b> 若您使用本数据集，请引用对应的发表论文： 针对MEDISEG（3-Pills与32-Pills）： bibtex @article{MEDISEG2026, title = {A dataset of medication images with instance segmentation masks for preventing adverse drug events}, author = {Chu, Wai Ip and Hirani, Shashi and Tarroni, Giacomo and Li, Ling}, journal={arXiv preprint arXiv:2603.10825}, year = {2026}, doi={10.48550/arXiv.2603.10825} } 针对MEDISEG-Deploy： bibtex @article{MEDISEGFewShot2026, title = {Evaluating Few-Shot Pill Recognition Under Visual Domain Shift}, author = {Chu, Wai Ip and Tarroni, Giacomo and Li, Ling}, journal={arXiv preprint arXiv:2603.10833}, year = {2026}, doi={10.48550/arXiv.2603.10833} } 若您的工作同时使用核心MEDISEG数据集与MEDISEG-Deploy，请同时引用两篇论文。