fiftyone

FiftyOne is an open-source Python library for dataset curation, labeling, visualization, and evaluation in machine learning, with a focus on computer vision. It provides a Python API and a lightweight web application that lets users explore datasets, annotate samples, and review model predictions.

Core concepts in FiftyOne revolve around datasets, samples, and fields. A dataset contains samples, each of which can store data and annotations in configurable fields. Supported label types include Detections (bounding boxes and confidences), Classification labels, Keypoints, Polylines, and binary or multi-class masks. Users can add custom fields to store metadata. Data can be imported from folders of images and from common annotation formats, and it can be exported to formats compatible with other tools.

Key features include interactive visualization and filtering in code or through the FiftyOne app, labeling workflows, and QA assistance. The library provides evaluation tools to compare predictions against ground truth and to compute metrics such as precision, recall, and mean average precision. It also offers dataset statistics, class distribution analyses, and the ability to view results across different subsets, making it useful for model analysis and dataset refinement. FiftyOne integrates with popular ML frameworks like PyTorch and TensorFlow, fitting into standard training and evaluation pipelines.

The FiftyOne app can be launched locally to browse datasets, inspect samples, and perform labeling tasks in a graphical interface. Programmatic usage involves creating a Dataset object, adding Sample entries with appropriate fields, and applying filters, views, and updates to prepare data for training or evaluation. It supports import/export to common formats and can be incorporated into existing machine learning workflows.

Since its release, FiftyOne has become a widely adopted tool for dataset management and model evaluation in computer vision, supporting both image and video data and ongoing community-driven development.