FiftyOne
Apache 2.0 Python library and app for curating vision datasets. Query and slice samples, explore embeddings, find label mistakes and push fixes back to CVAT or Label Studio.
Overview
Voxel51's answer to the question of what you do after the labels already exist. A Python library plus a browser app that treats a dataset as a queryable object: filter by confidence, sort by IoU against ground truth, group by class, then look at the actual images that come back. Apache 2.0 throughout, and since 2024 that includes the Brain module that used to be the withheld piece. Latest release 1.21.0, 19 August 2026, Python 3.10 through 3.13.
The Brain methods are the reason to bother. compute_similarity builds an embedding index and surfaces near duplicates, compute_uniqueness ranks which unlabelled samples are worth annotating next, and compute_mistakenness ranks existing labels by how likely they are to be wrong. Run that last one on a dataset somebody else annotated and the results tend to be uncomfortable. The evaluation API computes detection and segmentation metrics, then lets you click from a bad number straight into the samples producing it, which is a different activity from staring at a single mAP figure.
Curation here, drawing elsewhere, is the intended shape. Integrations create tasks in CVAT, Label Studio or Labelbox from a filtered view and merge the annotations back into the dataset. That handoff is not frictionless: attributes on a FiftyOne label can be lost when a round trip forces the label to be deleted and recreated, so schema discipline matters. The open source version is free and self-hosted with a MongoDB instance the pip package brings along. Enterprise is quoted per Voxel Processing Unit, each roughly 1,400 compute hours a month, with nothing published.
Key Features
- ✓ Dataset objects backed by MongoDB with a view language for filtering by label, confidence, field value or IoU against ground truth
- ✓ Brain methods: compute_similarity for near-duplicate detection, compute_uniqueness for sampling, compute_mistakenness for ranking suspect labels
- ✓ Model zoo including YOLO26, SAM3 with concept and visual video prompting, OneFormer and PP-OCRv6 text detection
- ✓ Evaluation API for detection and segmentation metrics with click-through from the metric to the failing samples
- ✓ Annotation integrations that create CVAT, Label Studio or Labelbox tasks from a filtered view and merge results back
- ✓ 3D and point cloud viewer, with Gaussian splat rendering and WebCodecs video decoding added in the 1.19 to 1.21 line
Where it holds
- • No open-source equivalent does dataset debugging at this depth. The nearest commercial comparisons are Encord and Scale's tooling
- • compute_mistakenness and compute_uniqueness surface bad labels and redundant samples rather than just plotting them prettily
- • Brain went fully Apache 2.0, so the interesting half is no longer withheld
- • Shipping fast: 1.19.0, 1.20.0 and 1.21.0 all landed between 14 July and 19 August 2026
Where it breaks
- • It labels nothing. You still need CVAT, Label Studio or a vendor for the actual drawing
- • The bundled MongoDB is not optional and is the usual culprit when a fresh install refuses to start on a hardened corporate machine
- • The CVAT round trip can silently drop label attributes when loading annotations forces a delete-and-recreate, and load_annotations has failed with more than ten CVAT projects present
- • Enterprise pricing is opaque. VPU-denominated quoting makes it hard to compare against a per-seat vendor without going through sales
My Take
First thing I would install on a new vision project. Load the dataset, run compute_similarity, and the near-duplicate clusters usually explain why the validation number looked suspiciously good. Version 1.21.0 arrived 19 August 2026 with Gaussian splat rendering in the 3D viewer and a rebuilt embeddings panel, which is a decent read on where the team is pointing. Two honest caveats: it draws no labels itself, and the bundled MongoDB is a real install hazard on locked-down machines.
Quick Info
- Pricing:
- open-source
- Openness:
- Open source
- Licence:
- Apache 2.0
- Starting at:
- FiftyOne open source is Apache 2.0 and free, including the Brain module which was closed until 2024. FiftyOne Enterprise carries no public list price and is quoted per deployment. Billing is denominated in Voxel Processing Units, where each VPU adds a Kubernetes pod to your cluster and provides roughly 1,400 compute hours a month. Verified on voxel51.com/pricing and voxel51.com/enterprise, August 2026.
- Added:
- Aug 2026
- Updated:
- Aug 2026
Use Cases
Judge it on your own work
The notes above say where FiftyOne holds and where it breaks. The fastest check is your own workload.
Visit website ↗Alternatives to FiftyOne
Roboflow
freemiumEnd-to-end computer vision platform for labelling, training and deploying detection models to cloud or edge hardware like Jetson, with a visual builder for multi-step vision pipelines.
CVAT
open-sourceMIT-licensed annotation tool for images, video and 3D point clouds. Self-host via Docker, plug SAM or YOLO into Nuclio for pre-labelling, export to COCO, YOLO, KITTI and 20 other formats.
Label Studio
open-sourceApache 2.0 labelling tool covering images, audio, text, video and time series in one config-driven UI. Self-host free, or $99 a month for HumanSignal's hosted Starter tier.