Diverse Dataset for Eyeglasses Detection: Extending the Flickr-Faces-HQ (FFHQ) Dataset

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Facial analysis is crucial for computer vision and machine learning applications.
• Existing datasets like FFHQ lack detailed annotations for facial accessories, specifically eyeglasses.
• Data-centric AI necessitates high-quality, diverse, and well-annotated datasets.

Purpose of the Study:

• To address the annotation gap in facial datasets for eyewear detection.
• To extend the FFHQ dataset with precise bounding box annotations for eyeglasses.
• To create a valuable resource for training and benchmarking eyewear detection models.

Main Methods:

• A semi-automated protocol was used to efficiently generate bounding box annotations for eyeglasses.
• The FFHQ dataset was extended to include over 16,000 images with eyewear.
• Deep learning models (YOLOv8, MobileNetV3) were used for baseline performance evaluation.

Main Results:

• The extended FFHQ dataset contains 70,000 images, surpassing CelebAMask-HQ in size and diversity.
• Models trained on the extended FFHQ dataset demonstrated superior performance in eyeglasses detection compared to those trained on CelebAMask-HQ.
• Cross-dataset validation confirmed the robustness of models trained on the enriched dataset.

Conclusions:

• The extended FFHQ dataset is a valuable resource for advancing eyewear detection research.
• The dataset supports data-centric AI approaches in facial analysis.
• Public availability of the dataset is expected to foster future research and development in facial accessory detection.