Efficient Face Detection in the Fisheye Image Domain.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Face detection in normal images has advanced significantly.
• Detecting faces in fisheye images is challenging due to severe peripheral distortion.

Purpose of the Study:

• To enhance face detection accuracy and speed in fisheye images.
• To develop a method that handles fisheye distortion without rectification.

Main Methods:

• A light-weight, location-aware network was proposed to differentiate central and peripheral image regions during feature learning.
• Anchor (bounding box) shape and scale were made location-dependent to suit the face detector.
• The system operates directly on fisheye images, independent of projection parameters.

Main Results:

• The proposed method demonstrated superior performance compared to the state-of-the-art RFB Net.
• Experiments were conducted on Wider-360 and real-world fisheye datasets.
• The system achieved high accuracy using only a single CPU core.

Conclusions:

• The developed location-aware network effectively addresses the challenges of face detection in fisheye images.
• The system offers a robust and efficient solution for real-time fisheye face detection.
• This approach eliminates the need for image rectification and calibration, simplifying the process.