tarsius

Local, training-free state detection for a region of a camera image: is the car in the garage, is the mailbox flag up, is the gate open.

You confirm a few frames per state, tarsius compares new frames against them and answers with a state or "unsure". Nothing leaves the machine and nothing gets trained.

from tarsius import Classifier, NcnnBackend, Roi, extract_features, load_image

backend = NcnnBackend()
spot = Roi(0.10, 0.10, 0.60, 0.50)
classifier = Classifier.for_backend(backend)
car = extract_features(backend, load_image(car_jpeg), spot)
empty = extract_features(backend, load_image(empty_jpeg), spot)
classifier.add("ref-1", car, state="occupied")
classifier.add("ref-2", empty, state="free")

result = classifier.classify(extract_features(backend, load_image(new_jpeg), spot))
result.status  # Status.CONFIDENT, Status.UNSURE or Status.NOT_READY
result.state  # "occupied" or "free" when confident
result.candidate  # best guess, also when unsure

Store features.reference().to_bytes() next to each confirmed frame and restore it with Reference.from_bytes(). On FeaturesError, extract again from the stored frame.

Backends

• NcnnBackend: SqueezeNet 1.1 ImageNet features on ncnn, install with tarsius[ncnn]. Tolerates clutter and lighting changes, a rough region is enough.
• EdgeBackend: numpy only, a baseline. Needs a tight region that still shows some static structure when the object is gone.

NcnnBackend is the intended default. It stays an optional extra until ncnn ships a release that no longer requires opencv-python, which has no musllinux wheels. Without that, tarsius cannot install on Alpine (musl) systems such as Home Assistant OS, whose Core container is Alpine based. ncnn master already made opencv optional (Tencent/ncnn@0ca5a4b), so the next ncnn release should unblock it.

Results so far

One household, two garage cameras, 24 days of recorded history, day and IR night. References come from older frames, the test is the newest week. The table is the harder camera: a dark car and clutter that gets moved around in front of it (459 frames, 164 in the test week). The other camera scores 100 % with both backends.

About 20 ms per 4 MP frame including JPEG decoding, for either backend, on an x86 desktop with one thread. A small sample, not a guarantee. scripts/history_split.py runs the same evaluation on your own recordings.

Development

scripts/build_model.sh   # once, builds the ncnn model into src/tarsius/models
uv sync --extra ncnn
uv run pytest
uvx prek run --all-files

The ncnn model is not stored in git. scripts/build_model.sh converts the original SqueezeNet 1.1 release with the ncnn tools. Inputs and outputs are checked against pinned sha256 sums, so CI and a local run produce the same files. It needs x86_64 Linux with curl, unzip, libgomp1 and libprotobuf32t64 (Ubuntu 24.04).

Releases: every PR gets a label, Release Drafter collects them into a draft, publishing the GitHub release builds and uploads to PyPI.

License

Apache-2.0. The bundled SqueezeNet model is BSD-2-Clause, see NOTICE.