FaceMe® AI Facial Recognition Engine | Tech Specs | CyberLink
FaceMe®

FaceMe® The World’s Top AI Facial Recognition Engine

FaceMe® Accuracy from NIST

National Institute of Standards and Technology (NIST)

NIST FRVT 1:1
VISA True Acceptance Rate
(@FAR 1E-4)
99.7%
VISA True Acceptance Rate
(@FAR 1E-6)
98.95%
VISA Border True Acceptance Rate
(@FAR: 1E-6)
99.06%
WILD True Acceptance Rate
(@FAR 1E-5)
96.88%

* CyberLink FaceMe® is ranked 16th globally and ranked 5th in NIST FRVT 1:1 on May 2020. if China & Russia vendors are excluded.

* VISA images are photos taken from the passport. The mean interocular distance (IOD) in the VISA images are 69 pixels.

* Wild images are non-constraint, photojournalism-style photos. Resolution varies widely. A wide yaw and pitch angle may be present.

* VISA Border are comparing VISA image to WebCam images.

FaceMe® Accuracy

FaceMe® Model
Ultra High Precision
(UH)
Very High Precision
(VH)
High Precision Model
(H)
Recommended Platform
Server, Workstation
PC, Premium Mobile
Mobile, Light-weight IoT/AIoT Devices
True Acceptance Rate
(@FAR 1E-4)
99.05%
98.95%
98.35%
True Acceptance Rate
(@FAR 1E-5)
99.05%
98.95%
97.03%
True Acceptance Rate
(@FAR 1E-6)
98.89%
97.99%
94.98%
Model Size (MB)
300 MB
17 MB
6.7 MB
Template Size (KB)
5 KB
3 KB
3 KB

Performance Numbers

For Workstation or Server

CPU
AMD EPYC 7502 32-Core
GPU/VPU
NVIDIA® RTX A6000
FaceMe® Extraction Model
UH
VH
Execution Time (ms) Detection + Extraction
3.9 ms
2.4 ms
Frames per second (fps)
256.4 fps
416.7 fps
Face Match Time (ms)
0.0018 ms
0.0017 ms
Memory (GB)
3.9 GB
3.7 GB

* 1080p images, 1 face per image

* The "High Precision (DNN)" model is used for Face Detection

* Face Match Time: Time taken to match two templates. For example, given a template extracted from a face image, to identify who is this person in a 1,000 people database takes 1,000 times matching

* FaceMe® Windows 5.4 is used

CPU
Intel® Xeon E3-1268L
GPU/VPU
NVIDIA® Tesla T4
FaceMe® Extraction Model
UH
VH
Execution Time (ms) Detection + Extraction
7.6 ms
5.2 ms
Frames per second (fps)
131.6 fps
192.3 fps
Face Match Time (ms)
0.0016 ms
0.0016 ms
Memory (GB)
1.2 GB
1.2 GB

* 1080p images, 1 face per image

* The "High Precision (DNN)" model is used for Face Detection

* Face Match Time: Time taken to match two templates. For example, given a template extracted from a face image, to identify who is this person in a 1,000 people database takes 1,000 times matching FaceMe Windows 5.1 is used

* FaceMe® Windows 5.1 is used

For Industrial PC

CPU
Intel® Celeron® G4920
Core i7-7700K
FaceMe® Extraction Model
VH
H
VH
H
Execution Time (ms)
Detection + Extraction
41.4 ms
24.9 ms
10.1 ms
8.1 ms
Frames per second (fps)
24.2 fps
40.2 fps
99.0 fps
123.5 fps
Face Match Time (ms)
0.0016 ms
0.0016 ms
0.0011 ms
0.0011 ms
CPU
Intel® Celeron® G4920
FaceMe® Extraction Model
VH
H
Execution Time (ms)
Detection + Extraction
41.4 ms
24.9 ms
Frames per second (fps)
24.2 fps
40.2 fps
Face Match Time (ms)
0.0016 ms
0.0016 ms
CPU
Intel® Core i7-7700K
FaceMe® Extraction Model
VH
H
Execution Time (ms)
Detection + Extraction
10.1 ms
8.1 ms
Frames per second (fps)
99.0 fps
123.5 fps
Face Match Time (ms)
0.0011 ms
0.0011 ms

* 1080p images, 1 face per image

* FaceMe® SDK - Windows 5.2 is used

* The "High Precision (DNN)" model is used for Face Detection

* For VPU case, test is done by running 960 images concurrently using Batch mode and measure the average time per image using FaceMe SDK 3.15

For Mobile Devices & IoT/ AIoT Devices

Soc
Qualcomm Snapdragon 845
Qualcomm Snapdragon 660 (GPU)
MediaTeK i350 (APU)
Apple A12X
Device
Google Pixel 3
Advantech MOD Q200
MediaTeK i350 Dev Kit
iPad Pro 12.9” 2018
FaceMe® Extraction Model
VH
H
VH
H
VH
H
VH
H
Execution Time (ms)
Detection + Extraction
55.1 ms
40.9 ms
62.3 ms
37.7 ms
123.2 ms
54.2 ms
29 ms
22 ms
Frames per second (fps)
18.1 fps
24.4 fps
16.0 fps
26.5 fps
8.1 fps
18.4 fps
34.5 fps
45.5 fps
Face Match Time (ms)
0.0006 ms
0.0006 ms
0.0010 ms
0.0010 ms
0.0035 ms
0.0035 ms
0.0025 ms
0.0009 ms
Soc
Qualcomm Snapdragon 845
Device
Google Pixel 3
FaceMe® Extraction Model
VH
H
Execution Time (ms) Detection + Extraction
55.1 ms
40.9 ms
Frames per second (fps)
18.1 fps
24.4 fps
Face Match Time (ms)
0.0006 ms
0.0006 ms
Soc
Qualcomm Snapdragon 660 (GPU)
Device
Advantech MOD Q200
FaceMe® Extraction Model
VH
H
Execution Time (ms) Detection + Extraction
62.3 ms
37.7 ms
Frames per second (fps)
16.0 fps
26.5 fps
Face Match Time (ms)
0.0010 ms
0.0010 ms
Soc
MediaTeK i350 (APU)
Device
MediaTeK i350 Dev Kit
FaceMe® Extraction Model
VH
H
Execution Time (ms) Detection + Extraction
123.2 ms
54.2 ms
Frames per second (fps)
8.1 fps
18.4 fps
Face Match Time (ms)
0.0035 ms
0.0035 ms
Soc
A12X
Device
iPad Pro 12.9” 2018
FaceMe® Extraction Model
VH
H
Execution Time (ms) Detection + Extraction
29 ms
22 ms
Frames per second (fps)
34.5 fps
45.5 fps
Face Match Time (ms)
0.0025 ms
0.0009 ms

* 720p images, 1 face per image

FaceMe® Technology

Features List
Face Detection
Yes
Face Recognition
Yes
Face Tracking
Yes
Face Attributes Recognition
Gender, Age, Emotion, Pose
Real-time Video Support
Yes
RTSP/H.264 Support
Yes
Emotion Detection
Happy, Sad, Angry, Surprised, Neutral
Images Pre-Processing
Lighting Enhancement, Edge Sharpening, Upsampling
Anti-Spoofing with 3D-Cameras
Yes
Anti-Spoofing with 2D-Cameras
(with user interactions)
Yes
Anti-Spoofing with 2D-Cameras
(without user interactions)
Yes
Chipsets (CPU/GPU/VPU)
Intel®64, ARM64, NVIDIA® GPU, NVIDIA® Jetson, Intel®Movidius VPU, MediaTek i350 APU, Qualcomm SoC(GPU), NXP i.MX8
Programming Interface
C++, HTTP, C#, Perl
AI Inference Engines
Intel® OpenVINO, NVIDIA® TensorRT / CUDA, Qualcomm SNPE, MediaTeK NeuroPilot, CoreML

Images/ Cameras Conditions

Minimum
Recommended
Face Pose – Yaw
< 60°
< 45°
Face Pose – Pitch
< 50°
< 30°
Face Pose – Roll
< 60°
< 45°
Lighting
450 lux
550 lux
Face Size
(pixels between eyes)
36 pixels
50 pixels
Face Enrollment
1 image
5 images - Straight, Left/Right (15°~45°), Up (5°~30°), Down (15°~30°)

Anti-spoofing using 3D-Depth Camera

Supported 3D-Depth Cameras
Intel® RealSense, Orbbec® Astra Pro, Himax®, Altek®, eYs3D®, iPad Pro
True Positive Rate
98.2%
True Negative Rate
100%

Fast Search Algorithm for an Enormous Database

People in Database
Time to Search in Database (ms)
One-by-one Compare
Fast Search
1 Million People
1,860 ms
1.2 ms
6 Million People
11,200 ms
1.6 ms

* With Fast Search algorithm, FaceMe can search a face within a 6 million person database for about 1.6 ms.

* Tested on a i7-7700K with 24GB memory PC. A database of 30 million faces would require 128GB memory.

Accuracy of Face Attributes Detection

Detection Attributes
Accuracy Rate
Gender
98%
Emotion
Up to 86%
Age
5.8 y/o (Mean Average Error)

* Tested with 68,000 images.

For further information on CyberLink FaceMe® engine