Multiview face detection and registration requiring minimal manual intervention.

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Traditional face recognition systems necessitate pre-detection and localization.
• Handling multiple faces with varying views and scales presents a significant challenge.

Purpose of the Study:

• To propose a novel approach for simultaneous detection and localization of multiple faces with arbitrary views and scales.
• To introduce a face constellation method that minimizes manual training data requirements.

Main Methods:

• Development of a face constellation for multiview face detection and localization.
• Automatic registration of training images to a reference image using distinctive local features.
• A probabilistic classifier for evaluating face correspondences and identifying distinctive points amidst false matches.

Main Results:

• The proposed face constellation method achieves accurate multiview face detection and localization.
• Minimal manual intervention is required for training, unlike other multiview approaches.
• Experimental results on FERET, CMU, and FDDB datasets demonstrate superior performance over state-of-the-art methods for arbitrary pose face detection.

Conclusions:

• The face constellation approach offers an efficient and effective solution for multiview face detection and localization.
• This method significantly reduces the need for extensive labeled training data.
• The approach shows strong performance in detecting faces across diverse and challenging poses.