Krill

Labeling tool for building and validating WAUV's RoboSub YOLO datasets, built to make every labeler as fast as possible.

Features

• Keyboard-first labeling. Label videos using only the keyboard
• SAM tracking. Segment Anything (SAM) model integration for automated tracked labeling
• YOLO pre-labeling. Train YOLO on verified labels right from the app.
• Model trials. Run any model over a clip or a whole video without touching the labels. Compare its boxes against the verified ones, see precision and recall per class, and jump to the frames it got most wrong.
• Review and accuracy. Reviewers check labelers' work, and automated checks ensure that labelers are accurate.
• YOLO exports. Preview and download a dataset split by video, with eval test videos kept out of training.
• Admin panel. Admin panel to manage users, content and monitor server health

Quick start

cp .env.example .env
just setup
just dev      # Postgres :5432, MinIO :9000 (console :9001)
just api      # :8080
just web      # :5173
# after registering, make your account an admin
just user-role you@example.com admin

To run the full stack with the GPU worker, set the passwords and KRILL_WORKER_TOKEN in .env, then:

just up
# after registering, make your account an admin
docker compose --env-file .env -f deploy/docker-compose.yml exec api krill user role you@example.com admin

The app is served from :8080. Browsers load frames straight from MinIO, so KRILL_S3_PUBLIC_ENDPOINT must be an address they can reach. After the first admin, roles are changed from the admin panel.

Contributing

Read CONTRIBUTING.md.

License

Copyright © 2026 Wisconsin Autonomous Underwater Vehicles, Camden Rush, and contributors.

Code released under the MIT License.