voxel51/fiftyone

What it solves

FiftyOne is designed to help developers build high-quality computer vision datasets and models. It addresses the challenge of visualizing and managing large-scale visual data, allowing users to identify data issues, fix annotation errors, and evaluate model performance more efficiently than manual inspection.

How it works

FiftyOne is an open-source tool that provides a powerful visual interface (the FiftyOne App) to explore images, videos, and their associated labels. It integrates with popular deep learning libraries like PyTorch and Hugging Face, allowing users to visualize embeddings, curate data, and perform quality assurance on 2D and 3D labels.

Who it’s for

It is primarily for computer vision engineers and researchers who need to organize, visualize, and evaluate their datasets and models.

Highlights

• Native Annotation: Create and edit 2D and 3D labels directly in the app or integrate with external tools.
• Visual Exploration: Explore complex datasets of images and videos with associated labels.
• Embedding Exploration: Select points of interest within embeddings to view corresponding samples.
• Model Analysis: Evaluate model performance and identify failure modes to guide fine-tuning.
• Data Curation: Quickly find and fix data issues, annotation errors, and edge cases.
• Libraray Integrations: Works with PyTorch, Hugging Face, and Ultralytics.