Why I built it

I kept running into the same problem while working with video datasets: the videos live in one folder, the annotations live in a pile of CSVs and detection files somewhere else, and there is no simple way to actually look at them together. Checking whether a face track is correct, or whether the labels line up with the right person, usually meant writing a throwaway script every single time.

Annie is my answer to that. It is a small, local-first web app that puts a video dataset and its frame-wise annotations side by side in the browser, so you can explore, sanity-check, and correct them in one place. Everything runs on a single machine — no server, no upload, no cloud. The datasets I had in mind are CMU-MOSEI and First Impressions V2 (face detections plus derived face tracks), but Annie is dataset-agnostic and built around extensible data sources.

What you can do with it

See it in action

The home tab (above) is a plain landing page that summarizes each tab and points you at a sensible starting workflow: convert your videos for consistency, define a dataset, then browse and review.

Browse is where most of the work happens. Each row is one sample: the original clip on the left, a strip of frames with the face detections drawn on them in the middle, and a full rendered preview on the right. The colored chips carry the annotations — subset, per-frame counts, tracks, and the label values for that sample. You can thumbs-up / thumbs-down, leave a note, and push anything questionable into the Annotator.

Try it

From PyPI

uv pip install "annie[all]"
annie

That launches the UI at http://127.0.0.1:8080. You will need FFmpeg (4–8) on your machine for frame decode and rendering — ffprobe ships with it.

With Docker

No local Python or FFmpeg required — the image on Docker Hub bundles everything.

docker pull fodorad/annie
docker compose up

Then open http://localhost:8080.

From source

git clone https://github.com/fodorad/Annie
cd Annie
make dev
make run