FiftyOne is an open-source platform, developed by Voxel51, for exploring, curating, and improving datasets used to train computer-vision and multimodal AI models. Its core is a Python library (installed with `pip install fiftyone`) built around a Dataset/Sample data model that stores images, videos, 3D/point-cloud data, and time-series alongside their metadata, labels, and model predictions. A browser-based FiftyOne App lets teams visualize and filter that data side-by-side with annotations and predictions in a single interactive view.
Beyond visualization, the FiftyOne Brain adds machine-learning-powered analytics for data-centric workflows: image and object embeddings for dimensionality reduction, similarity and duplicate search, uniqueness scoring to surface the most informative samples, and mistakenness detection to find likely annotation errors. Dedicated model-evaluation workflows let teams compare models, drill from aggregate metrics down to individual failing samples, and isolate edge cases, positioning FiftyOne as a curation-and-debugging layer rather than a labeling tool in its own right.
The project is designed to sit inside an existing ML stack via a plugin and operator framework and a broad set of integrations. Documented integrations span annotation tools (CVAT, Label Studio, Labelbox, V7), cloud storage (AWS S3, Google Cloud, Azure), vector databases (Qdrant, Pinecone, Redis, MongoDB), ML frameworks (PyTorch, TensorFlow), and model hubs (Hugging Face, Ultralytics), so datasets curated in FiftyOne can flow into labeling, similarity search, and training pipelines.
FiftyOne ships as a free Apache-2.0 open-source core, with a commercial FiftyOne Enterprise offering (marketed in Team, Growth, and Custom tiers) that adds collaboration, role-based access control, dataset versioning, and flexible deployment options including cloud, hybrid, on-premise, and air-gapped setups; Enterprise pricing is not published and is quoted via sales. By default the open-source package emits UUID-based import and App usage analytics, which can be disabled with the `do_not_track` config field or the `FIFTYONE_DO_NOT_TRACK` environment variable.
