Upload
satish-naidu
View
237
Download
0
Embed Size (px)
Citation preview
7/29/2019 Face Sparse MRF
1/28
Face Recognition With Contiguous
Occlusion Using Markov Random Fields
Zihan Zhou, Andrew Wagner, Hossein Mobahi, John Wright, Yi Ma
2010/11/181 zw
7/29/2019 Face Sparse MRF
2/28
Outline
Introduction
Local spatial continuity
Error correction
Choosing parameters
Experiments
2010/11/18 2
7/29/2019 Face Sparse MRF
3/28
Introduction
From ICCV09poster
Extended work from:
face recognition via sparse representation
Aimed at solving continuous error
Apply continuity constraint to sparse
representation
2010/11/18 3
7/29/2019 Face Sparse MRF
4/28
Introduction
J. Wright, A. Yang, A. Ganesh, S. Sastry, and Y. Ma.Robust face recognition via sparse representation. PAMI,
2009.
Sparse representation:
2010/11/18 4
y Ax
y Ax
=[ ]x
7/29/2019 Face Sparse MRF
5/28
Introduction
2010/11/18 5
y Ax e Solve this problem by:
7/29/2019 Face Sparse MRF
6/28
Introduction
2010/11/18 6
However, they only achieve their best performanceon occlusions that are not spatially correlated.
7/29/2019 Face Sparse MRF
7/28
Introduction
2010/11/18 7
7/29/2019 Face Sparse MRF
8/28
Introduction
2010/11/18 8
7/29/2019 Face Sparse MRF
9/28
Local Spatial Continuity
2010/11/18 10
support vector
7/29/2019 Face Sparse MRF
10/28
Local Spatial Continuity
2010/11/18 11
7/29/2019 Face Sparse MRF
11/28
Local Spatial Continuity
2010/11/18 12
Approximation to the log-likelihood function
7/29/2019 Face Sparse MRF
12/28
Error correction
2010/11/18 13
7/29/2019 Face Sparse MRF
13/28
Error correction
2010/11/18 14
1Estimating Linear Regressorx with Sparsity.
2Estimating Error Support s with MRF.
7/29/2019 Face Sparse MRF
14/28
Error correction
2010/11/18 15
3Identify the test image
assign it to the class that minimizes the error.1
l
7/29/2019 Face Sparse MRF
15/282010/11/18 16
7/29/2019 Face Sparse MRF
16/28
Choosing parameters
Choosing :
the level of error we would accept before
considering an entry of the image as occluded
2010/11/18 17
7/29/2019 Face Sparse MRF
17/28
Choosing parameters
Choosing :
2010/11/18 18
7/29/2019 Face Sparse MRF
18/28
Choosing parameters
Choosing :
2010/11/18 19
7/29/2019 Face Sparse MRF
19/28
Choosing parameters
Choosing :
controls the strength of mutual interaction
between adjacent pixels
2010/11/18 20
7/29/2019 Face Sparse MRF
20/28
Experiments
Recognition with synthetic occlusion
2010/11/18 21
7/29/2019 Face Sparse MRF
21/28
Recognition with synthetic occlusion
2010/11/18 22
7/29/2019 Face Sparse MRF
22/28
Experiments
Recognition with disguises
2010/11/18 23
7/29/2019 Face Sparse MRF
23/28
Recognition with disguises
2010/11/18 24
7/29/2019 Face Sparse MRF
24/28
Experiments
Validation on Extend Yale B dataset
2010/11/18 25
If
exceeds a thresholdit is declared as invalid
7/29/2019 Face Sparse MRF
25/28
First 19 subjects as training, the other 19subjects as invalid test images to be
rejected. IReceiver operating characteristic
(ROC) curves for 60% and 80% occlusion
Validation on Extend Yale B dataset
2010/11/18 26
7/29/2019 Face Sparse MRF
26/28
Experiments
Experiments with realistic test images
2010/11/18 27
Training: 116 subjects with 38 illuminations.
ITesting:
855 images under realistic illuminations (indoors,outdoors),
which have been divided into five categories:
Normal: 354 images
Occlusion by eyeglasses: 118 images
Occlusion by sunglasses: 126 imagesOcclusion by hats: 40 images
Occlusion by various disguises: 217 images
7/29/2019 Face Sparse MRF
27/28
Validation on Extend Yale B dataset
2010/11/18 28
7/29/2019 Face Sparse MRF
28/28
Validation on Extend Yale B dataset
29