RefineFace: Refinement Neural Network for High Performance Face Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Face detection technology has advanced significantly.
• Detecting small or tiny faces remains a considerable challenge.

Purpose of the Study:

• To introduce RefineFace, a high-performance, single-shot face detector.
• To address the challenge of detecting tiny faces with improved accuracy and recall.

Main Methods:

• RefineFace employs five key modules: selective two-step regression (STR), selective two-step classification (STC), scale-aware margin loss (SML), feature supervision module (FSM), and receptive field enhancement (RFE).
• STR enhances location accuracy by refining anchor initialization.
• STC, SML, FSM, and RFE collectively improve classification efficiency and discriminative feature learning, particularly for faces in extreme poses.

Main Results:

• RefineFace achieves state-of-the-art results on benchmark datasets including WIDER FACE, AFW, PASCAL Face, FDDB, and MAFA.
• The method demonstrates high performance in detecting tiny faces.
• RefineFace operates at 37.3 FPS using a ResNet-18 backbone for VGA-resolution images.

Conclusions:

• RefineFace offers a robust solution for high-performance face detection, particularly excelling in scenarios with numerous small faces.
• The proposed modular approach effectively enhances both localization and classification capabilities.
• The detector provides a favorable balance between accuracy and speed.