The boring work of converting CV dataset formats

What it does

Datumaro handles the unglamorous work of moving computer vision datasets between formats. It reads and writes more than a dozen annotation formats—including COCO, YOLO, PASCAL VOC, CVAT, and KITTI—and lets you merge, split, filter, and convert between them programmatically or from the command line. It also handles dataset hygiene, offering annotation validation, error checking, and statistics like image mean and standard deviation.

The interesting bit

The real utility isn’t just format shifting; it’s task-aware dataset surgery. Datumaro can split data while preserving label distributions for specific tasks—like ensuring no person ID leaks across train and test sets in re-identification work—or filter out occluded annotations and undersized bounding boxes without hand-rolled scripts.

Key highlights

• Bidirectional conversion among COCO, YOLO, PASCAL VOC, CVAT, KITTI, ImageNet, and other formats.
• Task-aware splitting that preserves label distributions for classification, detection, and re-identification.
• Built-in filtering for common cleanup jobs: dropping occluded labels, stripping small bounding boxes, or keeping only vertical images.
• Annotation validation and quality checks against task types, plus comparison against model inference.
• Dataset statistics including image-level mean and standard deviation.

Verdict

Worth a look if you manage CV datasets that cross format boundaries or need surgical filtering before training. If your data pipeline is already static and homogenous, it is probably overkill.