// the find
voxel51/fiftyone
Refine high-quality datasets and visual AI models
FiftyOne is a Python toolkit for visualizing, exploring, and cleaning computer vision datasets — think a local web UI where you can filter, tag, and inspect images/videos alongside their labels. It's aimed at ML practitioners who are drowning in unlabeled or mislabeled data and need something better than print statements and matplotlib. Actively maintained by Voxel51, which also sells an enterprise hosted version.
The embeddings panel is genuinely useful — you can run a model, project to 2D, and lasso a cluster of outliers to inspect them, which is faster than any manual alternative. Built-in support for uniqueness and mistakenness detection (via fiftyone-brain) catches labeling errors that would otherwise silently tank model metrics. The plugin system is real: you can write custom panels and operators that show up in the App without forking the project. Dataset zoo covers COCO, Open Images, and 50+ others with one-liner loading, which saves meaningful setup time.
It ships its own bundled MongoDB, which is cute for local use but a recurring source of complaints when versions conflict or the database gets corrupted — there is no clean migration path if something goes wrong with that embedded DB. The TypeScript/React frontend is a full monorepo with Recoil, Relay, and a custom command bus; contributing to the UI is a non-trivial investment, not a quick PR. Large video datasets are slow: it streams frames through Python, and you will feel it above a few hundred long clips. The open-source version quietly excludes collaborative features and cloud dataset connectors, which are gated behind Enterprise pricing that is not published anywhere.