Siamese Local and Global Networks for Robust Face Tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Convolutional Neural Networks (CNNs) excel in face detection, alignment, and recognition.
• Face tracking is vital for applications like video surveillance and human-computer interaction.
• Existing CNN-based face tracking methods are limited.

Purpose of the Study:

• To propose a novel Siamese CNN-based method for face tracking.
• To capture discriminative face information at both local and global levels.
• To introduce a large-scale dataset for evaluating face tracking methods.

Main Methods:

• Utilizing Siamese CNNs with hierarchical features learned from extensive face image data.
• Learning local representations for attribute patches (eyes, nose, mouth) for robustness against pose and occlusion.
• Learning global representations considering spatial relationships and facial characteristics (skin color, nevus).

Main Results:

• The proposed method effectively captures discriminative local and global face information.
• Experiments demonstrate superior performance compared to state-of-the-art visual tracking methods.
• A new large-scale dataset was created to benchmark face tracking.

Conclusions:

• The Siamese CNN approach offers an effective solution for face tracking.
• The method's ability to handle local and global features enhances tracking accuracy.
• The new dataset will advance research in challenging face tracking scenarios.