datumaro

Why we included this project

Datumaro is a Python library and CLI for anyone who has to move computer vision data between annotation formats. It converts between the formats teams actually use, COCO, PASCAL VOC, ImageNet, KITTI, Cityscapes, and CVAT exports among them, so you can drop the one-off conversion scripts. It also handles the tedious prep work that usually eats a day before training: merging datasets, filtering on custom criteria, splitting into train/val/test with task-aware splits that keep label and attribute distributions intact, and running quality checks against model inference. Annotation-level operations like turning polygons into instance masks or renaming labels make it a natural fit for annotation pipelines and dataset curation. If you maintain labeled CV datasets, especially ones feeding tools like CVAT, it gives you a consistent, reproducible way to keep that data in order.