synthetic data for face recognition

9
Synthetic Data for Face Recognition CS525 Vijay Iyer

Upload: reagan-lucas

Post on 01-Jan-2016

17 views

Category:

Documents


1 download

DESCRIPTION

Synthetic Data for Face Recognition. CS525 Vijay Iyer. Face Databases. Current databases (CMU PIE, FRGC/FRVT, FERET) Short range Indoors Artificial Light Only one known attempt at creating long range outdoor database CMU PIE small but very controlled dataset - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Synthetic Data for Face Recognition

Synthetic Data for Face Recognition

CS525Vijay Iyer

Page 2: Synthetic Data for Face Recognition

Face DatabasesCurrent databases (CMU PIE,

FRGC/FRVT, FERET)Short rangeIndoorsArtificial Light

Only one known attempt at creating long range outdoor database

CMU PIE small but very controlled datasetFERET , FRGC/FRVT large but sacrifice controlWe need more databases to further face

recognition

Page 3: Synthetic Data for Face Recognition

Why Synthetic?• Long term cost is cheaper(still costly

so this is not a deciding factor)• More experimental control• Explore more conditions • Can also be used to validate changes

in systems

Page 4: Synthetic Data for Face Recognition

4D Photohead FrameworkCustom Display Software

Allows for simple scripted animation3D Models

Generate models from CMU PIECreated with Animetrics Forensica software

Custom Display HardwareHigh power projector (3000 lumens)Cover blocks out light to improve visibility

Page 5: Synthetic Data for Face Recognition

4D Photohead Software

Page 6: Synthetic Data for Face Recognition

Model Validation

100%

Animetrics FaceGen

47.76%

Page 7: Synthetic Data for Face Recognition

Capture/Display Hardware

Page 8: Synthetic Data for Face Recognition

Initial ResultsDataSet Iso Distance V1 Comm.

FRGC Screen Shots N/A N/A 42.11 -FaceGenScreenShots N/A N/A 47.76 -AnimetricsScreenShots N/A N/A 100 -

PIE-3D-20100210B 500 81M 100 -

PIE-3D-20100224A 125 214M 58.82 100

PIE-3D-20100224B 125 214M 45.59 100

PIE-3D-20100224C 250 214M 81.82 100

PIE-3D-20100224D 400 214M 79.1 100Securics-1-02242010 125 214M 20 100Securics-2-02242010 250 214M 33.33 100Securics-3-02242010 400 214M 30 100

Page 9: Synthetic Data for Face Recognition

Summary/ConclusionsCreated an end to end framework which is

validated to work with frontal posesScientifically validated that the models facing

forward are equivalent to human beings for ROI of face recognition

Shown how synthetic data takes out or controls many existing variables in facial recognition.

Recent publication in the upcoming AMFG workshop shows the biometric research community has interest in developing this technique further.