Wide aspect ratio matching for robust face detection.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Anchor-based methods are standard in face detection but struggle with extreme aspect ratio faces due to sampling failures.
• Current methods often fail to capture faces with unusual aspect ratios, limiting detection model robustness.

Purpose of the Study:

• To improve face detection performance by extending the sampling range of face aspect ratios.
• To address the limitations of existing anchor matching strategies for extreme aspect ratio faces.

Main Methods:

• Theoretical exploration of factors influencing maximum Intersection over Union (IoU) for faces.
• Anchor matching simulations to evaluate aspect ratio sampling ranges.
• Proposal of a Wide Aspect Ratio Matching (WARM) strategy for collecting diverse positive anchors.
• Introduction of a Receptive Field Diversity (RFD) module for feature enhancement.

Main Results:

• The WARM strategy effectively collects more representative positive anchors across a wide range of aspect ratios.
• The RFD module provides diverse receptive fields, aiding in the capture of varied aspect ratios.
• Achieved promising Average Precisions (APs) on the WIDER FACE dataset: 0.965 (easy), 0.955 (medium), and 0.904 (hard).
• Demonstrated impressive generalization capability on the FDDB dataset.

Conclusions:

• The proposed WARM strategy and RFD module significantly enhance the ability of face detectors to capture extreme aspect ratio faces.
• The method offers a robust solution for improving face detection accuracy, particularly for challenging facial geometries.