# 🏆 **EdgeFire Grand Challenge 2025: Next-Gen AI for Early Fire Detection on Edge Devices**

---

## 🔥 **Overview**

The **EdgeFire Grand Challenge 2025** invites AI researchers, students, and Kaggle competitors to design **novel, efficient deep learning models** for real-time fire and smoke detection on **resource-constrained devices** such as Raspberry Pi, Jetson Nano, Android NPUs, and mini edge computers.

Participants will explore **lightweight architectures**, **multi-modal fusion**, **privacy-aware learning**, and **domain robustness** to build the next generation of **edge-deployable fire detection systems** capable of saving lives and forests worldwide.

This competition emphasizes:
- 🧠 Novelty in model design  
- ⚙️ Efficiency on edge hardware  
- 🧩 Robustness under real-world distortions  
- 🚨 Early detection capability  

---

## 💡 **Challenge Theme**

> “Design a novel, efficient, and explainable AI model for early fire and smoke detection that can run on low-power devices with high accuracy and low latency.”

---

## 🧠 **Objective**

Build and evaluate deep learning models capable of:
- Detecting **fire, smoke, and early onset cues**
- Performing well under **limited compute and memory**
- Maintaining **robustness** under lighting, compression, and weather distortions
- Supporting **quantization/pruning** for real-time edge inference

---

## 📊 **Dataset**

Participants will use the **Forest Fire Dataset for Binary Classification**, featuring:
- Two classes: `fire` and `nofire`
- Pre-split into **train**, **val**, and **test**
- Multi-source, with diverse lighting and weather conditions

They may **augment**, **expand**, or **restructure** the dataset (as long as their process is documented).  
Synthetic and weakly labeled data are allowed — provided they are **publicly reproducible**.

---

## 🧩 **Tracks**

### 🥇 **Track 1: Novel Architectures**

Implement or extend any of the following conceptual models — or propose your own:

| 🧠 **Model Name** | 🎯 **Concept Focus** | 💡 **Key Innovation** |
|--------------------|----------------------|------------------------|
| **FireFormer-E** | Early smoke cues | Temporal transformer with onset regression and haze-aware attention |
| **MultiFire-MM** | RGB + Thermal + Audio fusion | Tri-modal alignment with adaptive fusion and cross-modality voting |
| **SmokeSeg-YOLO-Decouple** | Lightweight detection + segmentation | Dual-head YOLO with shape-aware focal Tversky loss |
| **DR-FireNet** | Domain robustness | Frequency and style adapters for compression and lighting variation |
| **Fed-Private-Fire** | Privacy & scale | Federated LoRA adapters + DP-SGD for CCTV deployment |
| **EventFire** | Event-based low-latency detection | Spiking-inspired CNN for real-time grayscale/event data |

Participants can choose one, combine ideas, or propose an entirely **novel variant** — creativity is encouraged.

---

### 🥈 **Track 2: Edge Efficiency**

Deploy and benchmark your best-performing model on **edge hardware simulators** or **real devices**:

**Target Devices:**  
- Raspberry Pi 5  
- Jetson Nano  
- Jetson Orin Nano  
- Android NPU  
- CPU-only MiniPC  

Demonstrate model optimization via:
- Quantization (QAT or post-training)
- Pruning
- TensorRT / TFLite conversion

**You must submit:**
- 🧠 The compressed model file  
- ⚙️ An inference benchmark report (FPS, latency, power)  
- 📉 The model size and FLOPs count  

---

## ⚙️ **Evaluation Metrics**

| 🧮 **Metric** | 📖 **Description** | 🏋️ **Weight** |
|----------------|-------------------|---------------|
| **Accuracy / mAP** | Fire vs. NoFire detection accuracy | **30%** |
| **Latency (ms/frame)** | Inference time on Raspberry Pi / Jetson | **25%** |
| **Model Size (MB)** | Smaller is better | **15%** |
| **Novelty & Technical Soundness** | Innovation in architecture or optimization | **15%** |
| **Explainability & Reporting** | Visual outputs, overlays, interpretability | **10%** |
| **Robustness** | Performance under JPEG, fog, night, or blur | **5%** |

---

## 🧪 **Submission Requirements**

Each submission must include:
- 📔 **Notebook (.ipynb)** — Training + inference workflow  
- 🧠 **Model weights** — `.tflite`, `.onnx`, or `.pt` (max 100 MB)  
- 📄 **Technical report (PDF)** — Short 2–4 pages explaining:  
  - Architecture overview  
  - Optimization techniques used  
  - Hardware benchmarks  
  - Visual examples / explainability results  
- ⚡ **Inference log** showing model speed and power usage  
- 🎥 *(Optional)* Edge hardware video demo (YouTube or Drive link)

---

## 📦 **Baseline & Benchmark**

A simple CNN baseline (e.g., **MobileNetV2 with 65–70% accuracy**) will be provided as a **Benchmark Submission**.  
Your model must outperform this benchmark in at least **one dimension** (accuracy, latency, or model size).

---

## 🏁 **Evaluation & Leaderboard**

- **Public Leaderboard:** 40% test data (Public split from `solution.csv`)  
- **Private Leaderboard:** Remaining 60% test data (revealed post-competition)  
- 🏆 Winners determined by **Weighted Composite Score** (accuracy + efficiency + novelty)

---

## 🧾 **Evaluation Environment**

- **OS:** Ubuntu 22.04 / Jetson Linux / Raspberry Pi OS  
- **Frameworks:** PyTorch, TensorFlow, or ONNX  
- **Evaluation Tools:**  
  - `onnxruntime`  
  - `tflite_runtime`  
  - `torch.profiler`  
  - Power measurement via `tegrastats` or simulated runtime  

---

## 🌱 **Encouraged Techniques**

- Quantization-aware training  
- Pruning and distillation  
- Cross-modal alignment (RGB + thermal or audio)  
- Frequency-domain robustness  
- Federated privacy adaptation  
- Spiking or temporal transformers  

---

## 🧭 **Explainability & Trust**

Models must provide interpretable outputs, including:
- 🔥 Smoke and flame overlays  
- 🎯 Grad-CAM or feature visualizations  
- 📈 Uncertainty or confidence maps  

**Bonus points for:**
- False-alarm review loop  
- Operator dashboard mockup  

---

## 📜 **Rules Summary**

- Teams up to **5 members**  
- Original work only (no pre-trained commercial APIs)  
- Code and model weights must be **reproducible**  
- External data allowed if **publicly accessible** and **credited**  
- Winning submissions released under **MIT License**

---