Billion-Scale Face Recognition

Pose- and occlusion-invariant face recognition behind production identity at scale.

A line of work on face recognition that stays robust under the conditions that break naive 2D pipelines: large pose variation, partial occlusion, expression, and dataset shift between training and deployment. The thread runs through my Ph.D. and shipped components for production identity systems handling billion-level annual checks.

The 3D-aided 2D face recognition system UR2D (Xu et al., 2017; Xu et al., 2017) uses a 3D model registered against the 2D image to handle pose variation up to 90°. Joint head pose estimation and face alignment (Xu & Kakadiaris, 2017) share global and local CNN features, which we extended into face reconstruction with proper feature aggregation (Xu et al., 2019). Under occlusions specifically, OREO (Xu et al., 2020) improved generalization by 10.17% rank-1 accuracy in single-image settings. The FaRE package (Xu & Kakadiaris, 2019) packaged consistent open-source evaluation across these benchmarks.

The thread connecting all of it: robustness, not raw accuracy, is the binding constraint of deployed face recognition. That lesson carried into the production identity work and the lifelong-learning research that followed.

References

2020

  1. On Improving the Generalization of Face Recognition in the Presence of Occlusions
    Xiang Xu, Nikolaos Sarafianos, and Ioannis A. Kakadiaris
    In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2020

2019

  1. On the Importance of Feature Aggregation for Face Reconstruction
    Xiang Xu, Ha Le, and Ioannis A. Kakadiaris
    In IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2019
  2. FaRE: Open Source Face Recognition Performance Evaluation Package
    Xiang Xu and Ioannis A. Kakadiaris
    In IEEE International Conference on Image Processing (ICIP), 2019

2017

  1. IJCB
    Evaluation of a 3D-Aided Pose-Invariant 2D Face Recognition System
    Xiang Xu, Ha Le, Pengfei Dou, and 2 more authors
    In International Joint Conference on Biometrics (IJCB), 2017
  2. arXiv
    When 3D-Aided 2D Face Recognition Meets Deep Learning: An Extended UR2D for Pose-Invariant Face Recognition
    Xiang Xu, Pengfei Dou, Ha A. Le, and 1 more author
    arXiv preprint arXiv:1709.06532, 2017
  3. FG
    Joint Head Pose Estimation and Face Alignment Framework Using Global and Local CNN Features
    Xiang Xu and Ioannis A. Kakadiaris
    In IEEE Conference on Automatic Face and Gesture Recognition (FG), 2017