cvsuite

cvsuite is a computer vision toolkit for reproducible, filesystem-to-filesystem dataset work, organized around one CLI:

cvsuite <branch> <src> [transform] <output>

Convert between annotation formats (YOLO, COCO, LabelMe, class folders, segmentation masks, VQA manifests). Auto-label raw images from a text prompt with foundation grounding models such as SAM 3 or Grounding DINO. Caption them or answer questions about them with vision-language models. Sort them into class folders with zero-shot classifiers. Read text off them with OCR. Generate and edit images with diffusion models. Then do the unglamorous half of the job: unzip and flatten a dump of files, drop exact and near-duplicates, fix orientation, normalize size and format, and draw a balanced sample for training.

cvsuite label ./dataset-yolo to-coco ./dataset-coco
cvsuite label ./images ground --provider gsam --prompt "fire, smoke" to-coco ./labeled
cvsuite vlm ./images caption --provider qwen to-json ./captions
cvsuite class ./images infer --provider clip --prompt "good,bad" to-class-dir ./sorted
cvsuite gen create --provider flux --prompt "a red fox in the snow" to-dst ./fox.png
cvsuite prep arrange ./dump ./clean --unzip --exact-dedup --rename-seq

More in the cookbook.

You point a branch at a source, optionally run a transform, and materialize a destination layout. There is deliberately no universal "convert anything to anything" command.

How it works

flowchart LR
    SRC["source<br/>yolo · coco · labelme<br/>images · shards · …"] --> TOOL

    TOOL{"pick a tool<br/>label · class · vlm · gen"} -->|ingest| DS["VisionDataset"]

    DS --> TF["one transform<br/>sample · ground · ocr<br/>infer · caption · vqa"]
    TF -->|"chain runs to stack"| DS

    DS -->|output command| DST["destination<br/>yolo · coco · class-dir<br/>json · shards · …"]

Whichever tool you pick, the spine is the same: the source is ingested into a shared VisionDataset, transforms mutate that dataset, and an output command materializes it. What differs per tool is the ingest heuristics, which transforms exist, and the export rules.

Each run applies at most one transform, so you stack them by chaining runs: the destination of one becomes the source of the next. FM-backed transforms (ground, ocr, infer, caption, vqa, and the gen actions) hand the dataset to an isolated provider subprocess that mutates records in place. cvsuite prep is the exception to all of this, since it works directly on folders and never builds a dataset.

Branches

Install

pip install cv-suite

Or from source:

git clone https://github.com/ptonso/cv-suite
pip install -e ./cv-suite

The base install ships the CLI and the I/O layer. Foundation-model backends run in isolated per-model virtualenvs provisioned on first use. See docs/getting-started.md.

Documentation

• docs/ contains the guides: the mental model, getting started, and a cookbook of common workflows.
• specs/ is the technical source of truth: architecture, the dataset contract, per-branch behavior, business decisions, and the guardrails that must not break.

Names

• PyPI distribution: cv-suite
• import package and CLI: cvsuite

License

MIT. See LICENSE.