The fotonet Model Family

fotonete

fotonete is the supported model ID. It is a lightweight NMS-free object detector designed for constrained hardware and real-time inference: 1.72M parameters, 5.42 GFLOPs at 640×640, and no post-processing stage beyond a threshold and a cap.

Measured locally on the project's own test bench. Accuracy numbers are on the benchmarks page, with full methodology.

Loading a model

Fotonet("fotonete") resolves the verified release checkpoint for you, downloading it once and caching it locally. There is nothing to install by hand and no weights in the Git repository.

from fotonet import Fotonet

model = Fotonet("fotonete")

results = model.predict("street.jpg", conf=0.25, imgsz=640)
for det in results[0].boxes:
    print(det.cls, det.conf, det.xyxy)

Two environment variables override resolution when you need to:

• FOTONETE_MODEL_PATH — use a specific local checkpoint file.
• FOTONETE_CACHE_DIR — change where downloaded weights are cached.

How it compares

Accuracy is the honest trade-off in this project. fotonet is built for speed and footprint first: at 1.72M parameters it is roughly a third the size of the peers it is measured against, and it runs NMS-free. It does not out-accuracy larger models today, and the benchmark page says so plainly rather than hiding it.