fotonet logo
fotonet Real-Time Object Detection

What fotonet sees

Every box below is a real prediction recorded from the models running on our test PC — nothing is hand-drawn.

Live Browser Inference

Run fotonete directly in your browser with ONNX Runtime Web. Inference stays on your device, with WebGPU acceleration and automatic WASM fallback.

0.25
Loading model... WebGPU / WASM 0.0 ms — FPS 0 detections
Detected objects fotonete (client-side)
No objects detected yet. Start webcam or upload media.
The test bench
One desktop PC, measured locally
How to read these numbers

Every figure on this page was measured on our own test bench in a clean process. All models were evaluated on the same machine and the same COCO val2017 sample through pycocotools. Timing, memory, and cold-start figures are averages of 15 independent runs. Absolute values remain rig-specific; use the methodology and provenance in the benchmark when comparing runs.

Benchmarks

Eight metrics across fotonete and the detectors we compare it against, ranked by the measured result.

Get started in Python

From installation to detections in a few lines — then the faster paths for production use.

from fotonet import Fotonet
import cv2

# Load the deployment weights (1.7M parameters)
model = Fotonet("fotonet_last.pt")

# Run detection on an image
frame = cv2.imread("image.jpg")
results = model.predict_bgr(frame, imgsz=640, conf=0.25)

for det in results[0].boxes:
    print(det.cls, det.conf, det.xyxy)
# High-throughput path: pinned memory + FP16, one reusable buffer
from fotonet import Fotonet, FastPredictor

model = Fotonet("fotonet_last.pt")
fast = FastPredictor(model, device="cuda", half=True)

# dets is an (N, 6) array: x1, y1, x2, y2, score, class
dets = fast.predict(frame)
from fotonet import Fotonet

model = Fotonet("fotonet_last.pt")

# Standard deployment formats
model.export(format="onnx", path="fotonete.onnx", imgsz=640)
model.export(format="engine", half=True)   # NVIDIA TensorRT FP16
model.export(format="coreml")              # Apple Neural Engine
# Train from scratch on your own dataset
fotonet train model=fotonete data=my_dataset.yaml epochs=300 batch=64 imgsz=640

# Resume an interrupted run from the latest snapshot
          fotonet train model=fotonete data=my_dataset.yaml resume=fotonet_last.pt
Documentation

Go deeper with the docs

Installation details, the architecture behind fotonete, export guides for TensorRT and ONNX, and how to train on your own data.

Open documentation

Full comparison table

Every row measured on the same PC, in the same session, with the same concurrent-training load.

Model Architecture Params GFLOPs E2E latency B1 GPU pass B8 GPU pass Peak GPU memory Cold start COCO mAP

• fotonete is still mid-training — the row shows its latest checkpoint, not a finished model. Peer rows are the fully trained weights each project releases.

• Peer models were run through their own official packages and source on this machine; no numbers are quoted from vendor pages.

• mAP is evaluated on COCO val2017 (the measured sample set, conf 0.01, max 100 detections per image) with identical settings for every model.