FiftyOne

FiftyOne is an open-source tool from Voxel51 for building high-quality datasets and computer vision models. It is installed as a Python package and includes an interactive App for exploring images, videos, and 3D data alongside labels, metadata, and model predictions. The repository is released under the Apache 2.0 license, and the latest release listed is v1.22.1.

What It Is

FiftyOne is a visual AI data workflow tool. Users load datasets from their own data or a dataset zoo, then visualize, curate, label, and evaluate them in one place. A quickstart loads a small zoo dataset and launches the App from a Python session.

Core Workflows

The documentation organizes the product around data import and export, pre-annotation curation, annotation, post-annotation curation, model training, inference, and evaluation. Curation features include embedding visualizations, similarity search, near-duplicate detection, and hard-sample mining. Annotation covers creating and editing 2D and 3D labels natively in the App or through connected annotation tools. Model evaluation covers classification, detection, segmentation, and regression, with the goal of identifying failure modes.

Extensibility and Ecosystem

A plugin system lets users extend the App with operators and panels, written in Python and TypeScript. The docs list integrations with annotation backends (CVAT, Label Studio, V7, Labelbox), vector databases (Qdrant, Pinecone, Milvus, LanceDB, and others), and ML libraries (Hugging Face, Ultralytics, PyTorch Hub, OpenCLIP). Dataset and model zoos are included. The documentation also describes agent-related features, including skills and a FiftyOne Agent in the Enterprise edition.

Setup Path

Install with pip install fiftyone on Python 3.10 to 3.13, ideally in a virtual environment. FFmpeg is needed for video datasets, and a source install requires Node.js and Yarn to build the App. Voxel51 also offers FiftyOne Enterprise for collaborative, cloud-native workloads.