pyronear/yolov11s

Pyronear YOLO model for early wildfire smoke detection.

Release name: Tenacious Tiger Latest version: v8.3.0

Each release is a git tag on this repo (e.g. v8.3.0). Pin a version with revision="v8.3.0" in hf_hub_download / snapshot_download.

Model details

Field	Value
Architecture	yolo11s
Image size	1024
Epochs	50
Optimizer	AdamW
Weights SHA-256	c4b5b846b957cef2...
Training data MD5	31c0e3708401fa0c...

Files

File	Description
best.pt	PyTorch weights
onnx_cpu.tar.gz	ONNX export (cpu)
ncnn_cpu.tar.gz	NCNN export (cpu)
manifest.yaml	Full training manifest

Usage

PyTorch (ultralytics)

from ultralytics import YOLO

model = YOLO("best.pt")
results = model.predict("image.jpg", imgsz=1024, conf=0.2, iou=0.01)
for r in results:
    print(r.boxes)  # bounding boxes + confidences

ONNX (onnxruntime)

from huggingface_hub import hf_hub_download
import onnxruntime as ort
import numpy as np
from PIL import Image

path = hf_hub_download(repo_id="pyronear/yolov11s", filename="onnx_cpu.tar.gz")
session = ort.InferenceSession(path, providers=["CPUExecutionProvider"])

img = Image.open("image.jpg").resize((1024, 1024))
x = np.array(img).transpose(2, 0, 1)[None].astype(np.float32) / 255.0
outputs = session.run(None, {session.get_inputs()[0].name: x})

NCNN

# Unzip first
tar -xzf ncnn_cpu.tar.gz

Download with huggingface_hub

from huggingface_hub import snapshot_download

# latest release
local_dir = snapshot_download(repo_id="pyronear/yolov11s")
# or pin a version
local_dir = snapshot_download(repo_id="pyronear/yolov11s", revision="v8.3.0")

Pyronear engine (sequential smoke detection)

from pyroengine.engine import Engine

engine = Engine(
    conf_thresh=0.20,
    nb_consecutive_frames=5,
)
# feed frames one by one — engine.predict() returns a score
score = engine.predict(pil_image, cam_id="camera_01")
if score > engine.conf_thresh:
    print("Smoke detected!")

About Pyronear

Pyronear builds open-source tools for early wildfire detection.

License

The model weights in this repository (PyTorch .pt, ONNX and NCNN exports) are released under the GNU Affero General Public License v3.0 (AGPL-3.0), see the LICENSE file.

They were trained with the Ultralytics YOLO framework (YOLO11, AGPL-3.0) starting from Ultralytics pretrained weights, and are therefore distributed under the same license. If you use these weights in a product or service, the AGPL-3.0 obligations apply (including making the corresponding source available to users), unless you hold an Ultralytics Enterprise License.