innomium/fire-detection

<div align="center">

<img src="innomium_icon.svg" width="88" alt="Innomium logo" />

Innomium Ember

Ultra-light YOLO fire, smoke, and extinguisher detection for safety monitoring and edge vision.

90% accuracy · 3 classes · 9.8 MB ONNX · Runs in browser & on edge

  

</div>

Overview

Innomium Ember is a cutting-edge YOLO-based fire hazard detector — very light, very strong, and built for real-world safety scenes.

Detect fire, smoke, and fire extinguisher with a compact ~9.8 MB ONNX model that runs on CPU, edge hardware, or directly in the browser.

Performance

Ember classifies fire hazards in outdoor and industrial scenes with per-class bounding boxes and confidence scores.

<div align="center">

</div>

What this demonstrates:

• —All three hazard classes in one scene: fire, smoke, and fire extinguisher
• —Per-class NMS without cross-class suppression
• —Flame detections up to 77% alongside smoke (36%) and extinguisher (31%)
• —Suitable for outdoor events, safety drills, warehouses, and industrial sites

Key Features

• —Multi-class YOLO — Fire, smoke, and fire extinguisher with class ID remapping from the ONNX head.
• —Per-class NMS — Hard NMS applied independently per class to preserve overlapping hazard types.
• —Smoke merge — Fragmented smoke boxes merge into coherent plumes for cleaner alerts.
• —Color-prior filters — Borderline fire and extinguisher detections are validated against expected pixel appearance.
• —Privacy-first demo — The live Space runs inference entirely in your browser. No frames are uploaded.

Use Cases

Live Demo

Try Ember directly in your browser:

[https://huggingface.co/spaces/innomium/fire-detection](https://huggingface.co/spaces/innomium/fire-detection)

1. 1.Open the Space
2. 2.Upload an image or click Load Example
3. 3.View detections with class labels and confidence scores

Repository Structure

├── index.html              # Marketing site + interactive demo
├── main.js / detector.js   # Browser inference (ONNX Runtime Web)
├── weights.onnx              # ONNX model weights (Git LFS)
├── app.py                    # Python FireDetector for batch / server use
├── example_input.png         # Sample input frame
├── example_output.png        # Sample detection output
└── innomium_icon.svg         # Innomium logo

Local Python Usage

pip install -r requirements.txt

from pathlib import Path
import cv2
from app import FireDetector

detector = FireDetector(Path("."))
image = cv2.imread("example_input.png")
boxes = detector.predict_image(image)
for box in boxes:
    print(detector.class_names[box.cls_id], box.conf)

Batch inference

results = detector.predict_batch([image], offset=0, n_keypoints=0)
for frame in results:
    print(frame.frame_id, len(frame.boxes))

Model Pipeline

1. 1.Letterbox preprocess — Resize and pad to model input (640×640)
2. 2.ONNX inference — YOLO multi-class fire hazard detection
3. 3.Class remap — Map model head order to [fire, smoke, fire extinguisher]
4. 4.Per-class NMS — Independent hard NMS per hazard class
5. 5.Smoke merge & fire suppress — Merge fragmented smoke; suppress nested fire duplicates
6. 6.Color filters — Validate borderline fire/extinguisher boxes against pixel appearance

About Innomium

Innomium builds cutting-edge computer vision models for mission-critical environments — where accuracy, latency, and deployability all matter.

Very light. Very strong. Built for safety.

License

Apache 2.0