HarpoonNet 1.2 - Advanced Drone Detection Model

📄 License: Non-commercial use only. 🔐 Commercial licenses available upon request. Contact: [email protected]

🛡️ Commercial Use Notice

⚠️ This model requires explicit permission for commercial use. ⚠️

✅ FREE for: Research, education, academic use, open-source projects
❌ REQUIRES LICENSE for: Commercial products, revenue-generating applications, proprietary systems
📧 Contact: [email protected] for commercial licensing

Website

check us out @ chiliadresearch.com!

Updates

July 8- fixed bug causing users to download wrong model (Harpoon 1.1) instead of Harpoon 1.2 with the new ConvNeXt backbone. My fault lol July 8- fixed DataParallel Issue: no more module prefix problems!

🎯 Model Description

HarpoonNet 1.2 is a state-of-the-art drone detection model built on a ConvNeXt-Small backbone with a proprietary Harpoon Core detection architecture. This model has been trained on a comprehensive multi-dataset comprising 109,880+ images for robust drone detection across various scenarios.

🏗️ Architecture

Backbone: ConvNeXt-Small (~50M parameters)
Detection Head: Harpoon Core (~4.7M parameters)
Total Parameters: ~54.7M
Input Size: 544x544 pixels
Output: Single-class detection (drone)
Anchors: 3 anchor boxes per grid cell
Feature Map: 17x17 grid (544/32=17)

📊 Model Performance

Training Dataset: 110,000+ multi-domain drone images
Validation Loss: 0.059270 (enhanced ConvNeXt training)
Inference Speed: ~60 FPS on modern GPU
Model Size: ~122MB (PyTorch ConvNeXt-Small)
[email protected]: 95%+ (superior accuracy)

🚀 Quick Start

Installation

pip install torch torchvision opencv-python pillow numpy

Load Model

import torch
from harpoon_modular import create_harpoon_net_12

# Load the HarpoonNet 1.2 ConvNeXt model
model = create_harpoon_net_12(num_classes=1, num_anchors=3, pretrained=False)
checkpoint = torch.load('pytorch_model.pth', map_location='cpu')

# Handle both full checkpoint and weights-only files
if 'model_state_dict' in checkpoint:
    model.load_state_dict(checkpoint['model_state_dict'])
else:
    model.load_state_dict(checkpoint)

model.eval()
print("🚀 HarpoonNet 1.2 ConvNeXt model loaded successfully!")

Inference

import cv2
import torch
from torchvision import transforms
from PIL import Image

def preprocess_image(image_path):
    # Load and preprocess image
    img = cv2.imread(image_path)
    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
    img = cv2.resize(img, (544, 544))  # Updated resolution
    
    # Convert to tensor
    transform = transforms.Compose([
        transforms.ToTensor(),
        transforms.Normalize(mean=[0.485, 0.456, 0.406], 
                           std=[0.229, 0.224, 0.225])
    ])
    
    img_tensor = transform(Image.fromarray(img)).unsqueeze(0)
    return img_tensor

# Run inference
with torch.no_grad():
    img_tensor = preprocess_image('drone_image.jpg')
    predictions = model(img_tensor)
    detections = model.decode_predictions(predictions, confidence_threshold=0.85)  # Higher threshold
    
    # Process detections
    for detection in detections[0]['boxes']:
        print(f"Drone detected at: {detection}")

🎮 Real-time Detection

The model comes with ready-to-use webcam detection scripts:

Basic Detection

python webcam_detection.py --model pytorch_model.pth --conf 0.85

Advanced Tracking

webcam_detection_harpoonnet12.py - Enhanced detection for HarpoonNet 1.2 (if you got some sort of gpu to handle the load- if not, just make your own webcam code ill make one thats more suitable at some point haha)

Controls:

q: Quit
+/-: Adjust confidence threshold
r: Reset tracker (tracking mode)
d: Toggle debug view

📁 Repository Contents

├── pytorch_model.pth          # Main model checkpoint
├── config.json               # Model configuration
├── training_history.json     # Training metrics and history
├── harpoon_modular.py        # Model architecture
├── config_multi_dataset.py   # Dataset configuration
├── LICENSE                   # Non-commercial license
└── README.md                 # This file

🔧 Model Configuration

Classes: 1 (drone)
Confidence Threshold: 0.85 (recommended for high precision)
NMS Threshold: 0.4
Input Resolution: 544x544
Normalization: ImageNet standard

📈 Training Details

Dataset Size: 109,880+ images from 8 datasets
Training Framework: PyTorch
Optimizer: AdamW with cosine annealing
Learning Rate: Enhanced warmup and decay
Augmentations: Advanced geometric and photometric
Validation Split: Stratified sampling
Best Epoch: 5 (validation loss: 0.059270)

🎯 Use Cases

✅ Permitted (Non-Commercial)

Academic Research: Computer vision studies and publications
Educational Projects: University coursework and learning
Open Source Projects: Non-profit community tools
Personal Experimentation: Hobby and learning projects

🔐 Requires Commercial License

Security Systems: Commercial perimeter monitoring
Airport Security: Professional UAV detection systems
Military Applications: Defense and surveillance contracts
Enterprise Software: Proprietary detection services
API Services: Commercial drone detection APIs

⚡ Performance Tips

GPU Acceleration: Use CUDA for optimal performance
Batch Processing: Process multiple images for efficiency
Confidence Tuning: Use 0.85+ for high precision applications
Input Quality: 544x544 resolution provides best accuracy
Lighting: Enhanced model performs well in various conditions

🛠️ Advanced Features

ConvNeXt-Small Architecture

Modern CNN Design: State-of-the-art computer vision backbone
Efficient Processing: Optimized for accuracy and speed
Robust Detection: Enhanced feature extraction capabilities

ByteTrack Integration

Persistent Tracking: Maintains object IDs across frames
Occlusion Handling: Robust to temporary occlusions
Motion Prediction: Kalman filter-based motion model
Track Management: Automatic track creation and deletion

Real-time Optimization

Enhanced Architecture: Improved speed-accuracy trade-off
Memory Management: Optimized memory footprint
Multiple Formats: PyTorch, ONNX, TensorRT support

🏢 Commercial Licensing

For commercial use, we offer flexible licensing options:

Enterprise License: Full commercial rights for internal use
OEM License: Integration into commercial products
API License: Commercial API service deployment
Custom Training: Specialized model training services

Contact: [email protected] for pricing and terms.

📝 Citation

If you use HarpoonNet 1.2 in your research, please cite:

@misc{harpoonnet2025,
  title={HarpoonNet 1.2: Advanced Drone Detection with ConvNeXt Architecture},
  author={Christian Khoury},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/christiankhoury05/harpoon-1-2}
}

📄 License

This model is released under a Custom Non-Commercial License.

✅ Non-commercial use: Freely permitted
❌ Commercial use: Requires explicit written permission
📧 Licensing: Contact [email protected]

See LICENSE file for complete terms.

🤝 Contributing

Contributions for non-commercial use are welcome! Please feel free to submit issues and enhancement requests.

📞 Contact

For questions, support, and commercial licensing:

Email: [email protected]
Website: chiliadresearch.com
GitHub: christiankhoury05
Hugging Face: christiankhoury05

🔄 Model Updates

v1.2: Current version with 109k+ dataset, ConvNeXt-Small backbone
v1.1: Previous version with EfficientNet-B0 backbone
v1.0: Initial release

christiankhoury05
/

harpoon-1-2