FiftyOne
Open-source toolkit for curating high-quality computer-vision datasets and finding model failure modes.
About
FiftyOne is the open-source standard for working with image and video datasets in computer vision: a Python SDK plus a visual App for exploring data, evaluating model predictions, finding annotation mistakes, and hunting the slices of data where models fail.
Its real power is in the workflows around dataset quality — similarity search, embeddings visualization, active-learning sample selection, and integrations with labeling and training pipelines. Voxel51 maintains it under Apache-2.0 with a commercial platform on top for teams. If your agent or pipeline touches vision data, FiftyOne is where you go to make sure the data deserves the model.
What it can do
- Visual dataset exploration for images and video at scale
- Model prediction evaluation with per-sample error analysis
- Embeddings, similarity search, and duplicate detection
- Active learning and hard-sample mining workflows
- Integrations with labeling tools and training frameworks
Setup
pip install fiftyone, load a dataset with fiftyone.zoo or your own loader, then launch fiftyone app open to explore it visually.
Prerequisites
Python 3.9+; image or video data to curate.
Use cases
Works with
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: 11,188 GitHub stars (voxel51/fiftyone); repo pushed 2026-10-15 (same-day activity); Apache-2.0 license confirmed in repo; maintained by Voxel51; officialUrl https://voxel51.com curl-verified 200; docs.voxel51.com verified 200; PyPI project verified 200. Manual review: App + Python SDK workflow, embeddings/similarity/active-learning claims verified against official docs. License: Apache-2.0.
★ 11,188 GitHub stars
Sources
Every fact above was checked against these sources. Where a claim comes from the provider and couldn't be independently verified, the text says so.