🔥A Label-Free Lightweight Prompt-Driven Cross-Modal Fire Detection on Robotic Edge Platforms

A lightweight, prompt-driven cross-modal learning framework for real-time wildfire monitoring on resource-constrained edge devices.

🌟 Highlights

• 90% accuracy on Kaggle Fire & Smoke dataset
• 2239 FPS inference on Raspberry Pi 5 (0.45ms latency)
• 0.36 MB model size with 8-bit quantization
• LoRA-based parameter-efficient fine-tuning
• Zero-shot capability for unseen environments

📊 Performance Comparison

Method	Acc. (%)	FPS	Latency (ms)	Size (MB)
COCA	64	10	99.81	1010
Few-shot LoRA	71	2062	0.48	0.46
Few-shot Hybrid	46	2062	0.48	0.46
GIT Caption	71	12	82.34	690
CLIP-ViT	91	273	3.65	6.70
P-LearnEdge (Ours)	90	2239	0.45	0.36

🚀 Quick Start

Installation

# Clone the repository
git@github.com:spilabkorea/plearnedge.git
cd plearnedge

# Create virtual environment
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

Dataset Setup

Download the Kaggle Fire and Smoke dataset:

# Download from Google Drive
wget --load-cookies /tmp/cookies.txt "https://drive.google.com/uc?export=download&id=1L_TOG_sWp4xI9ojwe3YHu46VxmCS5xP8" -O dataset.zip

# Extract
unzip dataset.zip -d data/

Expected structure:

data/
├── fire/
│   ├── 0000000001.jpg
│   └── ...
└── smoke/
├── 0000000001.jpg
└── ...

### Training
```bash
# Basic training

python model/plearnedge.py


### Inference
```bash
# Single image prediction + FPS
python model/plearnedge_fps.py

📖 Model Architecture

P-LearnEdge combines:

1. Lightweight CNN (4 conv layers, 128D embeddings)
2. LoRA layers (r=4, α=16) for efficient adaptation
3. CLIP-style dual-encoder architecture
4. Prompt-driven zero-shot classification

from models import FireClipModel

# Initialize model
model = FireClipModel(embedding_dim=128, num_classes=2)

# Inference
with torch.no_grad():
    logits = model(image)
    prediction = torch.argmax(logits, dim=1)

🔬 Reproducing Results

Baseline Comparisons

# CLIP-ViT baseline
python baselines/clip.py 

# COCA baseline
python baselines/coca.py

# Few-shot hybrid
python baselines/fewshot_hybrid.py

📄 Citation

If you use P-LearnEdge in your research, please cite:

@article{plearnedge2025,
  title={A Label-Free Lightweight Prompt-Driven Cross-Modal Fire Detection on Robotic Edge Platforms},
  author={HyeYoung et al.},
  conference={ICRCV},
  year={2025}
}

📝 License

This project is licensed under the MIT License - see LICENSE file.

📧 Contact

• Author: SPILab Research Team
• Email: support@spilab.kr
• Website: https://spilab.ai/

🙏 Acknowledgments

• Kaggle Fire and Smoke Dataset contributors
• OpenAI CLIP team for inspiration
• SPILab Corporation for support

---

Note: This is research code. For production deployment, please contact us for optimized versions.