展开查看详情
1 .Monday, February 1 , 2016 Face Recognition
2 .Motivation Overview of Methods Face Detection Face Alignment Face Representation Face Classifier Results
3 .Motivation: General Goal Goal 1: Given a picture of a person’s face Given a bag of possible names What’s the name of the person in the picture? Goal 2: Given two pictures of a person’s face Are these of the same person ?
4 .Motivation: General Goal Goal 1: Given a picture of a person’s face Given a bag of possible names What’s the name of the person in the picture? Goal 2: Given two pictures of a person’s face Are these of the same person ?
5 .Overview of Methods Face Detection Localize the face Face Alignment Factor out 3D transformation Feature Extraction Find compact representation Classification Answer the question
6 .Overview of Methods Face Detection Localize the face Face Alignment Factor out 3D transformation Feature Extraction Find compact representation Classification Answer the question
7 .Methods for Detection Cascaded Ada-boosting [ P Viola 01 ] Deep Neural Net [ M Osadchy 07 ]
8 .Methods for Detection Cascaded Ada-boosting [ P Viola 01 ] Deep Neural Net [ M Osadchy 07 ]
9 .Challenges in Face Alignment Infer 3D from 2D Slight occlusion Lighting condition Head orientation Non rigid deformation
10 .DeepFace Alignment: Substep 1 2D feature point extraction 2D alignment Only for in plane alignment Fiducial Point Detection 2D Transformation Until convergence
11 .DeepFace Alignment: Substep 2 3D feature point extraction 3D alignment: piecewise affine transformation No perspective correction Reference 3D Fiducial Point Location Detected 2D Fiducial Point Location Final Alignment
12 .DeepFace Alignment: Substep 2 3D feature point extraction 3D alignment: piecewise affine transformation No perspective correction Reference 3D Fiducial Point Location Detected 2D Fiducial Point Location Final Alignment
13 .Global Feature: The EigenFace an eigen problem The set of images A The dictionary D The representation W [Turk 1991]
14 .Global Feature: Dictionary Learning I don’t want negative features: Nonnegative Matrix Factorization I want less non-zero elements: Compressed Sensing
15 .Local Features [ I Atanasova 2010] Down Sample Local Binary Pattern Laplacian SIFT Pros: easy, fast to compute Cons: not expressive enough
16 .The DeepFace [ Yaniv Taigman 2014] Convolution+ Rectified Linear Max pooling Convolution+ Rectified Linear Locally Connected+ Rectified Linear Fully Connected
17 .The DeepFace [ Yaniv Taigman 2014] Convolution+ Rectified Linear Max pooling Convolution+ Rectified Linear Locally Connected+ Rectified Linear Fully Connected
18 .Classifier Same Person Task: Metric Learning: SVM
19 .Classifier Name of the Person Task:
20 .Classifier Name of the Person Task:
21 .The Biggest Dataset Ever The SFC Dataset From Facebook 800-1200 each, 4030 people, 4.4M in all The LFW Dataset 13323 photos of 5749 celebrities
22 .The Necessity of Deep Neural Net More samples, less error Shallower Neural Net, more error Small error increase in bigger data set
23 .Comparison No alignment: 87.9% Only 2D alignment: 94.3% Full alignment + DeepFace : >97%
24 .Still Challenging On YTF dataset, from Youtube videos Due to motion blur, view angles
热门城市
