Author: thor: Gonza nzalo lo Benito to gonzabenito - - PowerPoint PPT Presentation
Mast ster er The hesis is Dissertat ertation ion Comparing classical and deep approaches for face recognition in a smartgym application Author: thor: Gonza nzalo lo Benito to gonzabenito nzabenito@gma gmail.co l.com Advisors:
Author: thor: Gonza nzalo lo Benito to gonzabenito nzabenito@gma gmail.co l.com Advisors: Sergio io Escaler lera, a, UB and d CVC Josep ep Lladó dós, , UAB and CVC 1
2
1. 1. Foo
tball ll Club Club Barcelona’s smart smartgym ym project aims to develop a system to aid gymnasium ium training ing. 2. System should identif tify and tra track ck down wn at athlet letes es within the facility, allowing self configuration of the exercise machines according to the user and computing training times of each one of them. 3. The fac ace re reco cognit gnition method should perform with precision in real-time on reduced sets of subjects and samples per individual.
3
4
5
[1] - P. J. Phillips, S. Z. Der, P. J. Rauss, and O. Z. Der, FERET (face recognition technology) recognition algorithm development and test results. Army Research Laboratory Adelphi, MD, 1996. [2] - M. Grgic, K. Delac, and S. Grgic, “Scface – surveillance cameras face database,” Multimedia Tools and Applications, vol. 51, no. 3, pp. 863–879, 2011. [3] - B. Heisele, B. Weyrauch, V. Blanz, and J. Huang, “Component-based face recognition with 3d morphable models,” 2012 IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops, vol. 05, p. 85, 2004.
[1] FERET ET: 1. 1. 14126 14126 images. 2. 2. 1199 1199 subjects. 3. 3. ~12 ~12 samples per subject. 4. 4. 180 180° face rotation. 5. 5. First t big face database.
6
[2] SCface: e: 1. 1. 4160 4160 images. 2. 2. 130 130 subjects. 3. 3. 32 32 samples per subject. 4. 4. 180 180° face rotation. 5. Samples with survei eillan lance ce cameras [3] MIT-CBC BCL: L: 1. 1. 3240 3240 images. 2. 2. 10 10 subjects. 3. 3. 324 324 samples per subject. 4. 4. 0° to 34° face rotation. 5. 5. 3D synthetic ic masks.
Top Face recognition methods tested on Labeled Faces in the Wild.
7
[2] [1] [3] [4] [6] [5]
[1] - F. Schroff, D. Kalenichenko, and J. Philbin, “Facenet: A unified embedding for face recognition and clustering.” in CVPR. IEEE Computer Society, 2015. [2] - Y. Sun, D. Liang, X. Wang, and X. Tang, “Deepid3: Face recognition with very deep neural networks.” CoRR, vol. abs/1502.00873, 2015. [3] - O. M. Parkhi, A. Vedaldi, and A. Zisserman, “Deep face recognition.” in BMVC, X. Xie, M. W. Jones, and G. K. L. Tam, Eds. BMVA Press, 2015 . [4] - Y. Taigman, M. Yang, M. Ranzato, and L. Wolf, “DeepFace: Closing the Gap to Human-Level Performance in Face Verification,” in 2014 IEEE Conference on Computer Vision and Pattern Recognition. IEEE, Jun. 2014. [5] - X. Cao, D. P. Wipf, F. Wen, G. Duan, and J. Sun, “A practical transfer learning algorithm for face verification.” in ICCV. IEEE Computer Society, 2013. [6] - D. Chen, X. Cao, L. Wang, F. Wen, and J. Sun, “Bayesian face revisited: A joint formulation.” in ECCV (3), ser. Lecture Notes in Computer Science,
[1] - M. Turk and A. Pentland, “Eigenfaces for recognition,” J. Cognitive Neuroscience, vol. 3, no. 1, pp. 71–86, Jan. 1991. [2] - P. N. Belhumeur, J. P. Hespanha, and D. J. Kriegman, “Eigenfaces vs. fisherfaces: Recognition using class specific linear projection.” pp. 711–720, 1997. [3] - O. M. Parkhi, A. Vedaldi, and A. Zisserman, “Deep face recognition.” in BMVC, X. Xie, M. W. Jones, and G. K. L. Tam, Eds. BMVA Press, 2015, pp. 41.1–41.12.
8 Representative of classica cal l approa
ches.
Only CNN method with offici cial l code made public lic.
9 Eigenfaces seeks to ma maxi ximi mize ze da data ta scatter tter, and find the principal vectors to describe the image space. A new image x is project cted ed into the face space by: With uk an eigenface. This gives the descripto tor by the weight vector
=
Eigenface #1 Eigenface #2 Eigenface #3 Eigenface #4 Eigenface #5
+ . . .
mean face Face space
10 Fishe herfac faces es adds the condition
mizing the ratio of the determinant of the between ween- clas ass scatte tter matrix of projected samples, to to the determinant of the withi hin- class scatter tter matrix of the projected samples.
Max overall scatter Optimized scatter
1. Based on the CNN from 2014 2014 Simo monyan and Zisserman’s proposal. 2. 2. VGG VGG Face ce was trained specifically to identi dentify 2622 subj ubjects ects from a cus ustom tom built uilt database on celebrities’ face pictures [1]. 3. With triplet plet loss loss tr traini ning ng reaches 0.9913 accuracy
[1] - O. M. Parkhi, A. Vedaldi, and A. Zisserman, “Deep face recognition.” in BMVC, X. Xie,
11
Triplet training loss used in [1] VGG Face architecture
12 Hardware platform:
SDRAM and one NVIDIA IA GeFor
ce GT 940M 2GB DDR3. Program language:
enfaces ces and Fisher herfa faces ces programmed on Matla lab R2016b for Windows.
G Face programmed on Pycaff ffe for Linux. Dataset preparations: All samples were processed with fronta ntaliza lizatio tion, cropping the face detect ction
Original 20° Frontalized Cropped detection
MIT-CBCL BCL datab tabase used.
13
14
Face recognition with trai train and te test st se sets ts spl plit its over a database of frontaliz talized ed faces using a 10 10-folds
15 Performance comparison for each face rotation angle between the 3 models, when varying samples per subject. Eigenfaces in blues ues, Fisherfaces in greens eens, VGG Face in browns
16
epochs
▪ Reached near real-tim time recognition with over 95% accuracy (Eigenf nface ces and d Fishe herfa faces ces) for non optimized code, on MIT-CBCL CBCL datab tabase. ▪ Viable le field d tests ts with chosen methods. ▪ MIT-CBCL dataset taset is too small ll for VGG Face and facilitates overfittin ting without intense supervision vision. 17 Futur ure e work includes udes: ▪ Construction of FCB’s athletes face database. ▪ VGG Face tripl plet et loss traini ning to boost perfo forma mance nce. ▪ Tests ts on data with more subjects ects. ▪ Tests ts on simulta ltaneous neous recogniti gnitions ns to define final hardw dware are requirements.
18 Ex Experimen perimental tal setup etup for face recognition in FCB’s gy gym facilities based on Lin Linux ux and Rasp Raspber erry Pi Pi platform was developed. It includes the cameras, server and dedicated application to to build uild FCB’s database tabase.
Raspberry Pi unit
Application to build FCB’s gym database. Tests for face recognition with the setup in laboratory.
19
Aknowle nowledgement ements: To F.C. Barcelon Barcelona for their support and disposition to carry this collaboration project, in particular to Raul Peláez Blanco nco and Joan Ramón ón Tarragó gó. To Carlos los Báez and Coen Antens ens from CVC.