github.com/kitware/dive ↗

About kitware/dive

DIVE is a web-based platform for annotating and analyzing videos and images, built as a collaboration between Kitware and the VIAME computer vision project. The tool supports both manual annotation tasks—including bounding boxes, polygons, and lines on individual frames as well as multi-frame tracking with automatic interpolation—and automated analysis through integration with VIAME's machine learning algorithms. It handles most common video formats through automatic transcoding and allows users to define custom labeling schemes with text, numeric, and multiple-choice attributes.

The system is designed for deployment flexibility, available as a public web service at viame.kitware.com or as self-hosted Docker instances for local or cloud environments. The architecture uses Girder 5 as its data management backbone with supporting services including MongoDB, RabbitMQ, and Redis, while the client is built with Vue and the job runner uses Celery and Girder Worker. DIVE can operate in GPU-enabled mode for running computationally intensive VIAME pipelines and model training, or in CPU-only mode for lighter deployments.

The platform serves researchers and practitioners in domains like marine biology and general object detection who need to combine manual annotation workflows with automated computer vision analysis. It accepts video files or image sequences as input, optionally with existing annotation CSV files, and outputs detection results as CSV files alongside visual plots showing algorithmic results.