CVAT

CVAT (Computer Vision Annotation Tool) is an open-source web-based tool for annotating images and videos to create labeled datasets for computer vision models. It was originally developed by Intel and is maintained by the open-source community. The project is released under the Apache 2.0 license.

Key features include support for multiple annotation types such as bounding boxes, polygons, polylines, points, and pixel-level masks, as well as 3D cuboids for automotive tasks. It supports annotation for both still images and video sequences, with frame-by-frame labeling, interpolation between frames, and AI-assisted labeling to speed up the process. The tool supports collaborative workflows with projects and tasks, user roles, and review/approval steps.

Data formats: Annotations can be exported in common formats used by major frameworks, including COCO, Pascal VOC, YOLO, and others; CVAT also provides its own internal format for import/export and can be integrated into machine learning pipelines via APIs.

Deployment and architecture: CVAT is designed to run locally or in cloud environments using Docker Compose or Kubernetes. It uses PostgreSQL for metadata storage, Redis for task queuing, and S3-compatible storage or MinIO for assets. The backend is built with Django (Django REST Framework) and a web-based frontend, enabling a modular and scalable labeling workflow.

Licensing and community: CVAT is open-source software and welcomes community contributions. It is widely used in research and industry for creating labeled datasets across various computer vision tasks.