Datumaro

Dataset plumbing is Datumaro's specialty: the Intel Open Edge Platform project reads, writes, and converts computer vision datasets across dozens of formats, including COCO, YOLO, Pascal VOC, ImageNet, Cityscapes, CVAT, KITTI, Open Images, LabelMe, and MNIST, from either a command line interface or a Python API. Beyond conversion it merges datasets, filters annotations by class, size, or attributes, splits data into train, validation, and test subsets while preserving distributions, computes image and annotation statistics, and validates labels for task-specific quality problems, which makes it a standard companion to the CVAT annotation tool for format wrangling. Everything runs locally on CPU with a plain pip install, no GPU involved, and the MIT license keeps it free for commercial pipelines. Documentation lives on the Open Edge Platform site, and the project sits in the OpenVINO ecosystem, where it feeds curated data into training and optimization workflows. At roughly 700 GitHub stars it is a modest project by count, but it fills a niche nearly every vision team eventually hits.