A 3D Face Recognition Algorithm Directly Applied to Point Clouds.

Area of Science:

• Computer Vision
• Biometrics
• Machine Learning

Background:

• 3D face recognition using point clouds shows promise but struggles with limited 3D data and non-rigid facial structures.
• Extracting discriminative features directly from 3D point clouds remains a significant challenge.

Purpose of the Study:

• To develop a robust 3D face recognition system that overcomes data scarcity and improves feature extraction from point clouds.
• To enhance the accuracy and reliability of 3D face recognition in the presence of occlusions, pose variations, and expression changes.

Main Methods:

• A novel framework generates large-scale virtual 3D face scans using morphable models and Gaussian processes, guided by limited real data.
• A dual-branch network employing kernel point convolution (KPConv) is proposed for direct extraction of non-rigid facial features from point clouds.
• A local neighborhood adaptive feature learning module with context sampling enhances discriminative feature extraction through hierarchical downsampling.

Main Results:

• The proposed method effectively addresses the scarcity of 3D facial data through large-scale virtual scan generation.
• The dual-branch network successfully extracts discriminative non-rigid facial features from point clouds.
• Experiments on FRGC v2.0 and Bosphorus datasets validate the method's effectiveness and potential.

Conclusions:

• Guiding 3D face recognition with a small amount of real data combined with synthesized data is effective.
• The proposed approach demonstrates significant potential for improving 3D face recognition accuracy and robustness.