FiftyOne Cheat Sheet

1Install & Dataset1.21

pip install fiftyone import fiftyone as fo★
Bundles a local MongoDB. First launch initializes the database.
dataset = fo.Dataset("my-data", persistent=True)★
Persistent datasets survive process exit (stored in the DB). Without it, the dataset is deleted when you quit.
fo.list_datasets() · fo.load_dataset("my-data") · dataset.delete()★
Datasets are named & reloadable by name across sessions.
len(dataset) · dataset.first() · dataset.stats() · dataset.head()
Inspect size, peek at samples, see storage stats.

2Samples & label typesthe data model

s = fo.Sample(filepath="/data/img.jpg") dataset.add_sample(s)★
A Sample = a media file + arbitrary fields. Add fields (labels, metadata, tags) as attributes.
s["ground_truth"] = fo.Detections(detections=[ fo.Detection(label="cat", bounding_box=[0.1,0.1,0.3,0.4])])★
Bounding boxes are normalized [x, y, w, h] in [0,1]. A field can hold any label type.
fo.Classification(label=...) · fo.Segmentation(mask=...) · fo.Keypoints · fo.Polylines★
The label types — classifications, detections, masks, keypoints, polylines (and 3D). Store GT & predictions in different fields.
s.tags.append("train") · s["weather"] = "rain" · s.save()
Tags & custom fields for filtering. Call save() after editing a sample fetched from the DB.

3Import from diskcommon formats

dataset = fo.Dataset.from_dir( dataset_dir="/data", dataset_type=fo.types.COCODetectionDataset)★
Import from a standard layout. Types: COCODetectionDataset, YOLOv5Dataset, ImageClassificationDirectoryTree, VOCDetectionDataset, ...
dataset.add_images_dir("/data/images") # images only, no labels
Or add_images_patt with a glob. Then attach predictions programmatically.
dataset.compute_metadata() # width/height/size per sample
Populate media metadata for filtering (e.g. by resolution).

4The App & sessionsvisual browser

session = fo.launch_app(dataset)★
Opens the interactive App (browser or inline in a notebook) — browse media with labels/predictions overlaid.
session.view = my_view · session.dataset = other★
Drive the App from code: push a filtered view to inspect exactly the samples you're debugging.
session.selected # sample ids you clicked in the App
Read back the user's selection to script actions (tag, delete, export).

5Views & ViewFieldquery your data

from fiftyone import ViewField as F★
F references fields in expressions — like a query DSL over samples & labels.
view = (dataset .match(F("metadata.width") > 1000) .sort_by("uniqueness", reverse=True) .limit(100))★
Views are lazy, chainable, non-destructive — they never modify the dataset. Stages: match, sort_by, limit, skip, shuffle, take.
dataset.filter_labels("predictions", F("confidence") > 0.8)★
Filter within a label field (keep only high-confidence detections) vs match which filters whole samples.
dataset.match_tags("val") · dataset.exists("predictions") · select_fields(...)
Filter by tags, presence of a field, or trim fields. Iterate a view like a list of samples.

6Dataset & Model Zooready-made

import fiftyone.zoo as foz dataset = foz.load_zoo_dataset("quickstart")★
Curated datasets in one call (coco-2017, open-images-v7, ...). split=, max_samples= to subset.
model = foz.load_zoo_model("yolov8n-coco-torch") dataset.apply_model(model, label_field="predictions")★
Run a zoo model over the dataset and store its outputs as a label field — instant predictions to evaluate.
dataset.compute_embeddings(model, embeddings_field="emb")
Or extract embeddings for similarity/visualization (feeds the Brain).

7Evaluate predictionsGT vs preds

results = dataset.evaluate_detections( "predictions", gt_field="ground_truth", eval_key="eval")★
COCO-style mAP evaluation. evaluate_classifications, evaluate_segmentations for other tasks.
results.print_report() · results.mAP() · results.plot_confusion_matrix()★
Aggregate metrics & plots. Precision/recall per class in one report.
# eval_key writes per-sample TP/FP/FN — filter to the failures view = dataset.sort_by("eval_fp", reverse=True)★
The FiftyOne workflow: evaluation tags each detection/sample, so you can visually inspect exactly where the model fails in the App.

8Brain: embeddings & curationfiftyone.brain

import fiftyone.brain as fob fob.compute_visualization(dataset, brain_key="viz")★
Embed + reduce (UMAP/t-SNE) so the App shows a 2-D embeddings plot — lasso clusters to find structure & outliers.
fob.compute_uniqueness(dataset) · dataset.sort_by("uniqueness")★
Score how distinct each sample is — pick the most informative to label (or drop near-duplicates).
fob.compute_similarity(dataset, brain_key="sim") results.find_duplicates() # and sort/query by nearest neighbors★
Similarity index — find near-duplicates and do image/text search in the App.
fob.compute_mistakenness(dataset, "predictions", "ground_truth")★
Surface likely label errors by comparing confident predictions to GT — the fastest data-cleaning lever.

9Aggregations & exportstats & out

dataset.count_values("ground_truth.detections.label")★
Class distribution in one call. Also distinct, bounds, count, sum, histogram_values.
view.export(export_dir="/out", dataset_type=fo.types.YOLOv5Dataset)★
Export a view (your curated subset) to any supported format — the round-trip back to training.
dataset.tag_samples("to_relabel") · dataset.untag_labels(...)
Bulk-tag a view for annotation handoff or filtering.

!Common gotchasread before shipping

bounding boxes are normalized [x,y,w,h]
Top-left + width/height in [0,1], not pixels and not [x1,y1,x2,y2]. Converting wrong makes boxes vanish or explode.
persistent=True or it's gone
Non-persistent datasets are deleted at process exit. Set persistent=True (or you'll rebuild every run).
views don't modify the dataset
A view is a lazy query — edits go on the underlying samples. To persist a subset, clone() the view into a new dataset.
save() after editing a sample
Mutating a sample pulled from the DB needs sample.save() (or view.save()) to persist — a silent no-op otherwise.
match vs filter_labels
match keeps/drops whole samples; filter_labels trims labels within a field. Mixing them up is a common surprise.

FiftyOne cheat sheet