Transcript 下載/瀏覽Download

KinectFaceDB : A Kinect Database
for Face Recognition
Rui Min, Neslihan Kose, and Jean-Luc Dugelay, Fellow, IEEE
IEEE TRANSACTIONS ON SYSTEMS, MAN, AND
CYBERNETICS: SYSTEMS, VOL. 44, NO. 11,
NOVEMBER 2014
Presenter : Siang Wang
Advisor :
Dr. Yen - Ting Chen
Date :
2015.03.25
2
Outline
Introduction
Review Of 3-D Face Databases
Kinect Face Database
Benchmark Evaluation
Data Quality Assessment Of KinectFaceDB And
FRGC
Conclusion
3
Introduction
The emerging RGB-D cameras such as the
Kinect sensor have been successfully applied
to many 3-D based applications.
•
•
•
•
Efficiency
Low-cost
Ease of RGB-D mapping
Multimodal sensing
4
Introduction
The adoption of this powerful new sensor for
face recognition has been mostly overlooked
due to the lack of a standard database for
testing.
5
Introduction
Face recognition,
the least intrusive biometric technique
applied to commercial and law enforcement
applications.
6
Introduction
That face recognition methods exploiting 3D cues are more efficient than 2-D-based
methods.
7
Introduction
Main contribution
A complete multimodal face database-based on
the Kinect is built and thoroughly described.
Extensive experiments are conducted for the
benchmark evaluations of 2-D, 2.5-D, and 3-D
based recognition using standard face
recognition methods.
8
Introduction
Main contribution
Recognition results on both the KinectFaceDB
and the FRGC
9
Review Of 3-D Face Databases
Face database is being proposed for two
main purposes
To test face recognition algorithms robust to one
or multiple facial variations
To assist with the development of face
recognition algorithms using a specific data
modality
10
Review Of 3-D Face Databases
Comparison between the 3-D & 2-D Face
Database
In comparison to the large number of 2-D face
databases, the number of available 3-D face
databases is relatively small.
3-D face recognition algorithms are less
affected by the illumination changes than 2-D
methods
11
Kinect Face Databases
The structure and the acquisition environment
of the proposed KinectFaceDB
Acquisition Process
The postprocessing steps
The potential usages of the proposed
KinectFaceDB
12
Kinect Face Databases
Database Structure
Fifty-two volunteers
13
Kinect Face Databases
Four types of data modalities are captured for
each identity
2-D RGB image
2.5-D depth map
3-D point cloud
RGB-D video sequence
14
Kinect Face Databases
nine facial variations
15
Kinect Face Databases
A protocol to record the RGB-D video
sequences
16
Kinect Face Databases
Acquisition Environment
Controlled indoor environment
Kinect (in front of participant)
White board (behind participant)
LED lamp(in front of participant)
17
Kinect Face Databases
Acquisition Process
Summarization of 3-D imaging procedure of the
Kinect
RGB and Depth Imaging From the Kinect
Converting to 3-D Face Data
RGB-D Alignment for Face Data
18
Kinect Face Databases
RGB and Depth Imaging From the Kinect
19
Kinect Face Databases
Converting to 3-D Face Data
20
Kinect Face Databases
The RGB and depth images are sampled
separately from two different cameras with a
displacement between them
21
Kinect Face Databases
RGB-D Alignment for Face Data
22
Kinect Face Databases
23
Kinect Face Databases
PostProcessing
Noise Removal
Facial Landmarking
24
Kinect Face Databases
Noise Removal
The point is too far
The point is too close
The point is in shadow cast by projecter
The surface reflects poor IR light
25
Kinect Face Databases
Facial Landmarking
Six anchor on the face
Left eye center
Right eye center
Nose-tip
Left mouth corner
Right mouth corner
Chin
26
Kinect Face Databases
Potential Database Usages in Addition to Face
Recognition
Facial Demographic Analysis
Three-Dimensional Face Modeling
27
Benchmark Evaluation
For 2-D and 2.5-D
PCA（Principal Component Analysis）
LBP（Local Binary Patterns）
SIFT（Scale-Invariant Feature Transform）
LGBP（Local Gabor Binary Pattern）
28
Benchmark Evaluation
29
Benchmark Evaluation
30
Benchmark Evaluation
For 3-D
ICP (iterative closest point)
TPS (thin-plate spline)
31
Benchmark Evaluation
32
Benchmark Evaluation
Fusion of RGB and Depth Face Data
combine both the RGB (2-D) and the depth
(2.5-D) face information from the Kinect in face
recognition
33
Benchmark Evaluation
34
Benchmark Evaluation
35
Data Quality Assessment Of KinectFaceDB
And FRGC
visually observe the 3-D
data quality differences
between the Kinect and
a high-quality laser
scanner
36
Data Quality Assessment Of KinectFaceDB
And FRGC
the depth accuracy and the depth resolution
Data quality
evaluate the identification/ verification
differences of 2.5-D/3-D faces captured by
the Kinect and a high-quality laser scanner
37
Data Quality Assessment Of KinectFaceDB
And FRGC
38
Data Quality Assessment Of KinectFaceDB
And FRGC
Kinect’s merits for face recognition
works in real time
3-D data provides complementary information
to the 2-D data, and can be easily integrated for
multimodal face recognition
with streaming 3-D video of the Kinect
39
Conclusion
complete multimodal (including well-aligned
2-D, 2.5-D, and 3-Dface data)
KinectFaceDB supplies a standard medium
to fill the gap between traditional face
recognition and the emerging Kinect
technology.
40
Conclusion
design of new algorithms and new facial
descriptors for the low-quality 3-D data
how to efficiently combine different data
modalities (RGB, depth,and 3-D) so as to
maximize the exploitation of the Kinect
41
Thanks for your attention