Revisiting Depth-Based Face Recognition From a Quality Perspective

Zhenguo Hu, Qijun Zhao, Feng Liu; The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2019, pp. 0-0

Abstract


Face recognition using depth data has attracted increasing attention from both academia and industry in the past five years. Despite the large number of depth-based face recognition methods in the literature, high quality data are usually required for high recognition accuracy. In this paper, we measure the quality of 3D face data in terms of resolution and precision, and evaluate how the accuracy of three deep face recognition models varies on several benchmark databases as the facial depth data resolution changes from dense to sparse and as the precision changes from high to low. From the experimental results, several observations are made. (i) Given a high precision, a low resolution of 3K is sufficient to represent a 3D face; when the precision decreases, using higher resolutions can benefit face recognition, but the recognition accuracy becomes saturated as the resolution reaches 10K. (ii) Depth precision is more critical than resolution in depth-based face recognition, and a precision of 1mm is generally preferred as a good balance between accuracy and cost. (iii) The deep models trained with low-quality data perform more stable across data of different quality levels. We believe that these observations are beneficial for both depth sensor manufacturers and depth-based face recognition system developers.

Related Material


[pdf]
[bibtex]
@InProceedings{Hu_2019_CVPR_Workshops,
author = {Hu, Zhenguo and Zhao, Qijun and Liu, Feng},
title = {Revisiting Depth-Based Face Recognition From a Quality Perspective},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
month = {June},
year = {2019}
}