voxel51/fiftyone

Refine high-quality datasets and visual AI models

FiftyOne – Open‑source toolbox for visual AI development

What it is – FiftyOne is a Python library (and accompanying web‑app) that helps you build, explore, label, and evaluate computer‑vision datasets and models. It works with images, video, and 2‑/3‑D annotations and integrates with popular frameworks such as PyTorch, Hugging Face, and Ultralytics.

Why it matters – High‑quality data is the biggest bottleneck for vision projects. FiftyOne gives you a visual interface to inspect raw data, spot annotation errors, explore model predictions, and run analytics (e.g., similarity search on embeddings) without writing custom code. This speeds up data‑curation cycles and model debugging.

Key capabilities

  • Native annotation – Create and edit 2‑D and 3‑D labels directly in the FiftyOne App or connect to external labeling tools.
  • Rich visual exploration – Browse images or video frames with their associated labels, filter by metadata, and view sample‑level details.
  • Embedding inspection – Project high‑dimensional feature vectors (e.g., from a backbone network) into a 2‑D view and select points to see the underlying samples.
  • Model evaluation – Compute standard metrics, visualise failure modes, and compare multiple runs.
  • Data curation & “Brain” analytics – Run built‑in utilities to find duplicate images, outliers, or mislabeled samples.
  • Extensible integrations – Plug‑in system lets you add custom visualisations, loaders, or connect to cloud services (e.g., Voxel51 SaaS).

Typical workflow

  1. Load a dataset – Use fiftyone.zoo to download public datasets or point FiftyOne at a local folder of images/video.
  2. Launch the App – session = fo.launch_app(dataset) opens an interactive UI in your browser.
  3. Explore & label – Filter, search, and add/modify annotations directly.
  4. Run analytics – Apply FiftyOne Brain utilities or custom scripts to surface problematic samples.
  5. Evaluate models – Attach predictions, compute metrics, and iterate.

Installation – One‑liner via pip (Python 3.10‑3.13):

pip install fiftyone

For developers, the repo can be built from source (requires Node/Yarn for the UI) or run inside Docker.

Where to learn more –

Who uses it – Researchers, data‑science teams, and enterprises building object‑detection, segmentation, video‑action‑recognition, or any vision pipeline that needs robust dataset management.

All information above is taken directly from the project's README.

Related

  • Project
    visual-layer/fastdup

    An open-source tool for analyzing and cleaning large-scale image and video datasets by identifying duplicates, outliers, and mislabels.

  • Project
    LibreYOLO/libreyolo

    An MIT-licensed computer vision library that provides a unified API for over twenty vision tasks, including detection, segmentation, and pose estimation, with integrated training and export tools.

  • Project
    CVHub520/X-AnyLabeling

    X‑AnyLabeling is a cross‑platform desktop application for AI‑assisted annotation of images, video, text, and 3D point clouds. It bundles dozens of state‑of‑the‑art vision models (YOLO, SAM, Grounding‑DINO, etc.) and supports local or remote inference, offering tools for detection, segmentation, OCR, pose, VQA, and more, with export to common dataset formats.

  • Dispatch
    The State of Computer Vision at Hugging Face

    Hugging Face has expanded its ecosystem to support 8 core computer vision tasks, over 3,000 models, and 100+ datasets, integrating both Transformer and convolutional architectures.

  • Project
    apple-aiml-research/ml-cvnets

    A computer vision toolkit for training and evaluating standard and novel mobile and non-mobile models for tasks like classification, detection, and segmentation.