Skip to main content

Quickstart: label your first video

This walks you through ingesting a video, running auto-labeling, reviewing the results, training a YOLO26 detector, and deploying it. Total time: about 15 minutes for a 60-second clip.

The pages we'll touch: dashboard → datasets → workflows → deploy.

0. Bring up the stack​

git clone https://github.com/oldhero5/waldo.git
cd waldo
cp .env.example .env
docker compose up -d

Wait ~30 seconds for Postgres to migrate, then:

docker compose logs app | grep -A 2 "bootstrapped first admin"

Save the printed admin password — it's the only time it's shown.

1. Sign in & create a dataset​

Open http://localhost:8000. The login page accepts the bootstrap credentials.

Login

Once you're in, the dashboard greets you with workspace stats and a "next step" nudge. Head to Datasets in the sidebar.

Dashboard

Click + New Dataset and give it a name.

Datasets

2. Upload a video​

Drop a .mp4, .mov, or .mkv into the upload zone. The backend extracts metadata via FFmpeg, stores the file in MinIO, and queues frame extraction.

Upload

For batch uploads from the command line:

TOKEN=...   # JWT from POST /api/v1/auth/login
curl -X POST http://localhost:8000/api/v1/upload/batch \
  -H "Authorization: Bearer $TOKEN" \
  -F "project_id=$PROJECT_ID" \
  -F "files=@clip1.mp4" \
  -F "files=@clip2.mp4"

3. Start a labeling job​

From the dataset, click Auto-label. Provide either:

  • Text prompts (one per line, e.g. person, car, truck) — SAM 3.1 grounded by free text.
  • Visual prompts — drag a box around an example object in the first frame; SAM 3 finds visually similar objects across the video.

Pick a confidence threshold (default 0.5) and resolution (1008 is a good middle ground). Click Preview to test on a handful of frames; click Start labeling to commit.

The job streams progress back to the UI over WebSocket. You can switch to Review as soon as the first frames complete.

4. Review​

The Review canvas shows each frame with overlaid boxes. Accept, reject, edit, redraw — every action PATCHes back to the API.

Review

Scrolling through reviewed frames.

Keyboard shortcuts:

KeyAction
J / KPrev / next frame
SpaceToggle play
DDelete the highlighted box
Shift+dragDraw a new box
RReject the whole frame

Rejected frames are excluded from training exports.

5. Train​

Open Train for the job. Pick a YOLO26 variant (yolo26n for fastest, yolo26m for a sensible default, yolo26x for max accuracy), pick an augmentation preset, and click Start training.

Train

Live logs stream from the trainer worker. Loss + mAP charts auto-scroll. The trained weights register in the model registry automatically when the run finishes.

6. Deploy​

Open Deploy → Models and star your new model to mark it active. The default /predict/* endpoints will use it from the next request.

Deploy

Stepping through the Deploy tabs.

Try it from the Test tab (drag in an image), or hit the API directly:

curl -X POST http://localhost:8000/api/v1/predict/image \
  -H "Authorization: Bearer $TOKEN" \
  -F "file=@sample.jpg"

That's the round trip — raw footage to a deployed model in one session.

Where next​