voxel51/fiftyone

★ 10,702PythonAudience · dataComplexity · 3/5LicenseSetup · easy

Mindmap

mindmap
  root((repo))
    What it does
      Visual dataset browser
      Label quality checks
      Model evaluation
    Use Cases
      Find mislabeled images
      Remove duplicates
      Build cleaner datasets
    Tech Stack
      Python library
      Local web interface
      Docker option
    Editions
      Community free Apache
      Enterprise paid tier
    Support
      Discord community
      Documentation site

mindmap root((repo)) What it does Visual dataset browser Label quality checks Model evaluation Use Cases Find mislabeled images Remove duplicates Build cleaner datasets Tech Stack Python library Local web interface Docker option Editions Community free Apache Enterprise paid tier Support Discord community Documentation site

Click or tap to explore — scroll the page freely

Things people build with this

USE CASE 1

Browse thousands of labeled images visually to find mislabeled, duplicate, or low-quality samples before training a model

USE CASE 2

Compare your computer vision model's predictions side-by-side with ground truth labels to identify where it consistently fails

USE CASE 3

Write Python queries to filter and export a cleaner subset of a large image dataset for a focused retraining run

Tech stack

PythonDocker

Getting it running

Difficulty · easy Time to first run · 30min

Free to use, modify, and distribute for any purpose including commercial use under the Apache 2.0 license.

In plain English

FiftyOne is an open-source Python tool for working with image and video datasets used in computer vision and AI development. It gives you a visual interface to browse large collections of labeled images, spot problems in the data, compare model predictions against ground truth, and decide which samples need more attention before you train or retrain a model. The core idea is that the quality of a visual AI model depends heavily on the quality of the data behind it. FiftyOne makes it practical to find mislabeled images, locate duplicate or near-duplicate samples, identify where a model is making consistent mistakes, and build cleaner, more balanced datasets as a result. You work through Python code that loads your dataset into FiftyOne, then launch a local web interface to inspect and filter it visually. Installation is a single pip command. The tool supports Python 3.9 through 3.12 and works on Mac, Windows, and Linux. For developers who want to contribute or build from source, the repository includes install scripts that also set up the front-end app. A Docker image is available as well. The project is built by Voxel51. There is an open-source community edition and a paid enterprise tier aimed at teams that need cloud-native collaboration, larger scale, and production infrastructure. The community edition is free and released under the Apache 2.0 license. FiftyOne integrates with common machine learning frameworks and dataset formats used in computer vision work. The documentation site, linked from the repository, includes getting-started guides, tutorials, and a full API reference. There is an active Discord community and the project also publishes a blog and newsletter.

Copy-paste prompts

Prompt 1

Using FiftyOne in Python, show me how to load a folder of images with YOLO-format labels, launch the visual app, and filter down to only the samples where confidence is below 0.5.

Prompt 2

I trained an object detection model and want to find the images where it makes the most mistakes. How do I load my model's predictions into FiftyOne and visually compare them to ground truth?

Prompt 3

Show me a FiftyOne workflow for finding near-duplicate images in a dataset of 10,000 photos and exporting a deduplicated version.

Prompt 4

Using FiftyOne, how do I tag mislabeled images interactively in the UI and then export only the tagged samples as a new labeled dataset?

Prompt 5

Walk me through installing FiftyOne, loading a sample dataset from the FiftyOne Zoo, and running the visual interface for the first time.

Open on GitHub → Explain another repo

← voxel51 on gitmyhub — every repo by this author, as a profile.

Verify against the repo before relying on details.