FiftyOne for inspecting vision datasets and predictions

How it works

FiftyOne organizes samples and fields in a dataset and provides views for filtering and inspection. That lets an engineer examine errors visually instead of relying on aggregate scores.

What you need

Load a dataset with stable sample IDs and define where predictions and ground truth live in its schema. Keep enough metadata to slice errors by source or condition.

What to try first

Import ground truth and predictions for a small test set. Filter for high-confidence false positives and inspect them by class or camera source.

Slice the error set by camera, annotation source or collection batch. Concentration in one slice can indicate a data issue rather than a general weakness.

Where to look for failures

Mistakes in label import can look like model failures. Confirm class names, coordinates and sample IDs when integrating a new dataset.

Related Franklin pages

Follow one annotation from labeling to training and evaluation. Check that frame numbers, class IDs and coordinate conventions survive the handoff.

• Create image and video labeling tasks with review and export workflows. CVAT for computer vision annotation
• Train and evaluate pixel-labeling models using shared configurations and dataset interfaces. MMSegmentation for semantic segmentation experiments

For an overview of the types of predictions vision systems produce, see Franklin’s glossary. Computer vision

Project documentation

• Current project README
• README at the reviewed revision