README.md

metadata

license: cc-by-nc-4.0
language:
  - en
  - ko
tags:
  - object-detection
  - advertising
  - computer-vision
  - deep-learning
  - advertisement
  - call-to-action
datasets:
  - custom
metrics:
  - mAP
  - precision
  - recall
model-index:
  - name: MADUP Ad Action Element Detection
    results:
      - task:
          type: object-detection
          name: Object Detection
        dataset:
          type: custom
          name: MADUP Internal Advertisement Dataset
        metrics:
          - type: mAP
            value: 0.81
            name: mAP@50
          - type: precision
            value: 0.83
            name: Precision
          - type: recall
            value: 0.84
            name: Recall

MADUP Ad Action Element Detection Model

Overview

A state-of-the-art deep learning model for detecting call-to-action elements in advertising images, developed by MADUP and released as a contribution to the AI research community.

Performance

Korea Testing Certification (KOTCA) Certified Performance on internal dataset:

Metric	Score	Description
mAP@50	0.81	Industry-leading performance
Precision	0.83	High accuracy in detection
Recall	0.84	Excellent coverage of elements
Optimization Runs	211	Optuna hyperparameter tuning

Model Description

This model automatically detects and localizes call-to-action (CTA) elements in advertising images using advanced deep learning techniques. It has been optimized through extensive hyperparameter tuning to achieve superior performance in real-world advertising scenarios.

Key Features

• High Performance: Achieves 0.81 mAP@50 score
• Extensively Optimized: 211 Optuna hyperparameter optimization trials
• Production Ready: Suitable for real-world advertising analysis and automation
• Open Source: Released for research and educational purposes

Detectable Elements

The model can identify four types of advertising elements:

• Text: Advertising copy and call-to-action text
• Rectangle: Buttons and clickable areas
• Banner: Promotional banners
• Capsule: Pill-shaped UI elements

Quick Start

Installation

pip install -r requirements.txt

Basic Usage

from inference import AdElementDetector

# Initialize the detector
detector = AdElementDetector(model_path="best_model.pt")

# Run inference on an image
results = detector.predict("advertisement.jpg")

# Extract bounding boxes
boxes = detector.get_boxes(results)
print(f"Detected {len(boxes)} ad elements")

Command Line Interface

# Basic inference
python inference.py --image advertisement.jpg

# Save visualization
python inference.py --image advertisement.jpg --save output.jpg

# Adjust confidence threshold
python inference.py --image advertisement.jpg --conf 0.5

Batch Processing

from inference import AdElementDetector

detector = AdElementDetector()

# Process multiple images
image_paths = ["ad1.jpg", "ad2.jpg", "ad3.jpg"]
results = detector.predict_batch(image_paths, batch_size=8)

for img_path, result in zip(image_paths, results):
    boxes = detector.get_boxes(result)
    print(f"{img_path}: {len(boxes)} detections")

Technical Details

Architecture

• Base Model: Deep Learning Object Detection Architecture
• Input Size: 486x486 pixels
• Training Epochs: 120
• Optimization: 211 Optuna trials for hyperparameter tuning

Optimized Hyperparameters

imgsz: 486
lr0: 0.000515
dropout: 0.1
mixup: 0.15
copy_paste: 0.1
mosaic: 0.0

Model Files

├── best_model.pt       # Trained model weights (50MB)
├── inference.py        # Inference script
├── config.yaml        # Model configuration
├── data.yaml          # Dataset configuration
├── requirements.txt   # Python dependencies
├── LICENSE            # CC BY-NC 4.0 License
└── README.md         # Documentation

Requirements

• Python 3.8+
• PyTorch 1.8+
• CUDA-capable GPU (recommended)
• 4GB+ RAM

Python Dependencies

ultralytics
opencv-python
numpy

License

This model is released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license.

• Permitted: Research, education, personal projects
• Not Permitted: Commercial use without explicit permission
• Commercial Inquiries: pr_chapter@madup.com

Contributing

MADUP is committed to advancing the field of advertising technology through open research. We welcome contributions from the community.

Public Contribution Goals

• Research Support: Enabling academic research and innovation
• Industry Advancement: Driving advertising technology forward
• Community Growth: Contributing to the open-source ecosystem

Citation

If you use this model in your research, please cite:

@misc{madup_ad_detection_2024,
  title={MADUP Ad Action Element Detection Model},
  author={MADUP Research Team},
  year={2024},
  publisher={HuggingFace},
  note={KOTCA Certified Performance - mAP@50: 0.81}
}

About MADUP

MADUP is a leading data-driven advertising technology company specializing in AI-powered advertising optimization solutions.

• Website: https://www.madup.com
• Research Team: pr_chapter@madup.com
• LinkedIn: MADUP Company

Support

• Bug Reports: Please use the Issues tab
• Feature Requests: Submit a Pull Request
• General Inquiries: pr_chapter@madup.com

---

Built with dedication by MADUP Research Team for the global AI community