Anchor Cascade for Efficient Face Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Face detection is crucial for facial analysis tasks like recognition and reenactment.
• Current methods, cascade and anchor-based detectors, have limitations in accuracy or computational efficiency.
• Anchor-based methods often require large, pre-trained neural networks, hindering efficiency.

Purpose of the Study:

• To develop an efficient anchor-based cascade framework for face detection.
• To enhance detection accuracy by incorporating contextual information.
• To create computationally efficient face detection models with high accuracy.

Main Methods:

• Devised an efficient anchor-based cascade framework named anchor cascade.
• Proposed a context pyramid maxout mechanism to leverage contextual information.
• Trained and evaluated models on FDDB and WIDER FACE benchmarks.

Main Results:

• Anchor cascade significantly improves detection accuracy compared to MTCNN (e.g., 0.9435 to 0.9704 on FDDB at 1k false positives).
• The framework achieves high accuracy with computationally efficient models.
• Maintained comparable inference speeds to existing methods.

Conclusions:

• The proposed anchor cascade framework is effective for accurate and efficient face detection.
• The context pyramid maxout mechanism enhances detection performance.
• Anchor cascade offers a practical solution for real-world facial analysis applications.