Overview

Tech5 Face recognition SDK is designed for integration of face recognition technology in third-party applications and systems and allows to detect, track and recognize faces on images of any source and quality. Additional features provide: image quality assessment, age and gender estimation, smart image compression.

Tech5 Face recognition SDK provides fast speed and very accurate recognition regardless race or nation, age, face expressions, partial occlusions and even face morphing.

The SDK is written on C++ and includes C++ headers and samples.

Supported OS: Linux 64-bit, Windows 7 or later (64-bit only), Android Lollipop or later (ARMv7, ARMv8), iOS 13 armv64

CPU and GPU are supported.

Features
- Face detection with landmarks
- Face tracking
- Face alignment
- Biometric template extraction
- 1:1, 1:N, M:N matching
- Age and gender recognition
- Image quality assessment
- Serialization
- Image smart compression
HW requirements
- Supported GPU:
  - architectures NVIDIA Pascal (GeForce GTX 1080Ti, 1080, 1070, 1060, Tesla P40, P4, P100, etc), Volta (Tesla V100, TITAN V),
  - Turing (GeForce RTX 2080Ti, 2080, 2070, 2060, Tesla T4, TITAN RTX)
- Supported CPU: SkyLake generation or later, ARMv7 (32-bit), ARMv8 (64-bit)
- Min RAM: For x86_64 - 6 Gb, for Android - 2 GB
- Recommended RAM: 16 Gb
Image requirements
- GIF, PNG, JPEG, TIFF RGB24 bpp
- Face size: at least 20 pixels between eyes

Important: Different versions of the engine (the builder module) extract not compatible biometric templates (FIRs), even if the SDK versions are compatible. The interfaces can be the same, but the system itself won't work after upgrade without re-extracting the biometric templates.

Face SDK Modules

Face Detector

Module detects faces on the image and returns array of objects, where every object is coordinates of the boundary, face landmarks and confidence parameter.
Confidence parameter is the internal assessment of the detection error possibility (where 0 - definitely detection error, 1 - definitely a face).
In case of finding several faces on the image, Face detector returns several objects.
Following use case scenarios are up to SDK user: user can send all detected objects for following recognition, or choose one of them.
See more

Face Selector

Module selects one face from all input faces based on one of three user selected algorithm:

Widest face on image
Biggest face confidence
Face closest to the image's center

Alignment

Module aligns an input image after detection stage for the following template extraction on Builder module.
Be aware that Alignment module may work incorrectly if you send an image with several faces on it.
See more

Builder

Module extracts biometric templates (FIRs) from images, received after alignment.
Builder module requires using of the same version of Fir Matcher!
Builder 100 requires Fir Matcher 100, Builder 101 requires Fir Matcher 101, etc.
Pay attention: Builder module has no internal assessment and extract template from any input, even if there is not face on the image.
Builder is supposed to be used after Detector or Detector&Alignment modules.
See more

Matcher

Module compares FIRs and returns the result, which is the coefficient equivalent -logFAR.
See more

Age&Gender

Module classifies age and gender based on the image received from alignment module.
See more

Quality checker

Module assesses image quality parametes (blur, exposure, rotations) and facial attributes (glasses, open and closed eyes, smile, occlusions).
See more

Face tracking

Module tracks faces in the video stream, grouping faces of the same person in one track.
Module notifies that a new face had appeared in the frame and has left the scene.
See more

Image compressor

Module for compressing and decompressing images to the minumum size with different compression levels.
See more

Getting started

Release Notes. 1.7. Mozart

Added iOS 13 armv64 support
Added new modules:
- image smart compressor
- quality assessment
Added new algorithms:
Matching module updates:
- added function of score conversion from (-logFAR) to persentage
- added 1:1 match using ID within template list
Base module updates:
- size of serialized objects is significantly reduced
face_sdk_utils updates:
- added --age_gender job
- added --detect_faces job
- added --quality job
- added --image_compressor job
- changed parameters of --request job
Removed algorithms:
- builder 102
- face_detector 100, 102
- liveness passive 100
- liveness cooperative
- tracker 100
Others:
- size of binary files for Android is significantly reduced
- removed deb and rpm packages dependency on boost package

Release Notes. 1.6. Mozart

Added new nns:
- builder 105, 106, 200,
- face_detectors 200, 201
- alignment 103
Added new modules:
- face selector
Base module updates:
- reduced size of biometric templates

Release Notes. 1.5. XUXU

Added support NUMA architecture for Fir Matcher
Added Face Detector version 103.
Added Alignment version 103.

Release Notes. 1.4. XUXU

Added M:N matching where M can be 1,2 and 4 images of the same person.
Added Face Detector version 102 (With landmarks prediction).
Added Builder version 104 (XUXU NN).
Added Liveness detection
Added Face tracking
Added support of NVIDIA Volta and Turing architectures
Improved alignment
All modules in one library now

Release Notes. 1.3. IrishCream beta4

Added Face Detector version 101 (Improved accuracy).
Added Builder version 102 (Irish Cream mobile NN) and 103 (Irish Cream Indonesian NN).
Added RPM packages for Centos 7.
Added DEB packages for Ubuntu 18.04.

Release Notes. 1.2. IrishCream beta3

Added Windows 7 or later (64-bit only).
Added Face Detector.
Added new Builder version 101 (Irish Cream NN).
Added new Alignment 102 with quality assessement feature.
Updated face_sdk_utils. Added alignment, resizing, quality assessement, building and matching features.
Updated fir_matcher, added support of multithread processing.

Thanks for choosing our SDK! If you have any questions, don't hesitate to contact us suppo.nosp@m.rt@t.nosp@m.ech5-.nosp@m.sa.c.nosp@m.om

Disclaimer