Post on 28-Dec-2015
transcript
My Project Title
-Sridhar Godavarthy
Contents• A Little Background: Blink• A Lot More Background: Strain as a Soft
Forensic Evidence– Facial Recognition– Culprits– Human anatomy as a feature– Strain Measurement
• Micro expression Detection using Strain Patterns– Challenges– Sample Strain patterns
• References
Contents• A Little Background: Blink• A Lot More Background: Strain as a Soft
Forensic Evidence– Facial Recognition– Culprits– Human anatomy as a feature– Strain Measurement
• Micro expression Detection using Strain Patterns– Challenges– Sample Strain patterns
• References
BLINK!!!
A Little Background
Introduction: Blink
• Why are some people brilliant decision makers?• How do some people act upon instincts?• Why are we unable to explain some decisions?
Blink Contd…
• Great decision makers are not ones that process the most information– Malcolm Gladwell’s ‘The statue that didn’t look
right’
• They are those who have perfected the art of “Thin Slicing”– Filtering out the very few factors that matter.
Navarasas – the Nine Emotions
Contents• A Little Background: Blink• A Lot More Background: Strain as a Soft
Forensic Evidence– Facial Recognition– Culprits– Human anatomy as a feature– Strain Measurement
• Micro expression Detection using Strain Patterns– Challenges– Sample Strain patterns
• References
Facial Strain Pattern as a Soft Forensic Evidence
V.Manohar, D.B.Goldgof, S.Sarkar,Y.Zhang
Some slides have been adapted from the Authors’ presentation
A Lot More Background
Facial Recognition• Face recognition has made huge advances– Picasa’s Web Albums– Sony’s “say cheese”( or is it CHEERS) detection
• “Almost” perfect– Picasa still confuses between closely related faces– Canon almost always never detects my face• Some say - might be because of my hair ;-)
• Has anyone used the Lenovo Face ID?• Because they use static images• Could be supplemented for better
performance.
Culprits (ICHE)• Illumination• Camouflage(Makeup/glasses)• Facial Hair• Expressions
• The Solution: Use methods based on Human Anatomy
Methods based on Human Anatomy
• Iris scan• Retina scan• Skull X-ray
• Disadvantage– Require Specialized equipment– Intrusive
• Proposed Alternative– Skin and tissues of the face
Elasticity
• Different materials have different elasticity• Elasticity can be modeled
strain
stressElasticity Known
Calculate
Authentic Author Slide
Facial Strain
• What is Facial Strain?– Strain on soft tissue when expressions are made.– Anatomical method– Uses a pair of frames to measure deformation
Facial Strain
• Why Facial Strain?– As it is a difference, it is independent of all the
earlier mentioned culprits(ICHE)
Facial Strain– ‘Visual Pattern’ is unique to
every face.• Easily quantifiable by
‘elasticity’• Hard to measure – non-
linear, inverse equations• Can be represented by strain
pattern under specific boundary conditions• Is unique to a person.
Measurement of Facial Strain
• Contact strain measurement equipment is already available.– Cannot be used if we are looking to identify
people at a Casino/Airport– Did I mention the actual applications of this paper• Soft forensics based on surveillance videos
Measurement of Facial Strain Contd…
• Two major steps1. Obtain motion field between two frames2. Compute strain image from above Motion field.
First Step – Obtaining Motion Field• Feature Based– Need to identify features – Difficult!– Features may be ill defined( when camouflaged)– Usually requires manual intervention– Produces a sparse motion field– Produce Good correspondence in large motion
• Optical Flow based– Fully automated– Dense Motion field.– Requires constant illumination
First Step – Optical Flow
• Observed motion over sequential image frames
Adapted Author Slide
Second Step – Strain Computation Type
• 3D Strain– Ideal– No high speed equipment available to capture
range images
• 2D Strain– Well – not much of a choice– Authors could use existing data.
Second Step – Optical Strain
• Variation of displacement values obtained from optical flow– Calculated by taking the derivative of each pixel
Sobel operator (central difference)
Authentic Author Slide
Strain Computation - methods• Finite Element Method– Forward modeling when Dirichlet condition is satisfied– Good at handling irregular shapes– Computationally expensive– This method is an approximation to the solution
• Finite Difference Method– Strain, a tensor, can be expressed derivatives of the
displacement vector– This can be approximated by a Finite Difference Method.– Very efficient when carried out on a regular grid.– This method is an approximation to the differential
equation
Finite Difference Method
• Finite Strain tensor
• Cauchy tensor
Integrating Strain Patterns
• Motion is mostly vertical– Strain pattern is dominated by its normal
components
• The strain magnitudes are scaled to gray levels– White = highest strain– Black = lowest strain
• It is now a pattern matching problem.
Review of Choices
• Motion field : Based on Optical flow• Strain Type: 2-D• Computation: Finite Difference Method
Examples
Identification and matching
• Strain Magnitude is now 1-D• Use PCA to perform matching
Experiments
• Experiments performed on– Normal light– Low light– Shadow light– Regular face– Camouflaged face– Frontal view– Profile view– Neutral expression– Open mouth
Experiments Contd…
• Subject may not perform the expression to the same extent every time– Experiments repeated on shorter, subsampled
videos
Results
• Strain measurement seems to be logically correct
• We do not discuss the PCA and hence the recognition results as they are outside the scope of this discussion.( But they were good)
• Acts as a supplement to existing recognition methods.
Contents• A Little Background: Blink• A Lot More Background: Strain as a Soft
Forensic Evidence– Facial Recognition– Culprits– Human anatomy as a feature– Strain Measurement
• Micro expression Detection using Strain Patterns– Challenges– Sample Strain patterns
• References
Micro expression Detection using Strain Patterns
Macro Vs Micro expressions
• Macro Expressions:– Large movement• Smile• Talking• Shaking head
• Micro expressions– Raising eyebrow– Fast blinking
Can you classify?
Where will it be used?
• Supplement lie detection– Very little noise
• As part of a general discussion• Bond might not have lost even the first time!
Ideal Frame Sequence
1 2 3 4 5 n6
a. 1-2 b. 3-4
c. 1-4
d. 1-3e. 4-6
f. 1-6
a 1-2 100
b 3-4 200
c 1-4 300
d 1-3 200
e 4-6 200
f 1-6 400
Strain Measurement for a Practical Frame Sequence
Macro Expression
Micro Expression
Noise
Challenges
• Small movements are inevitable• Macro expressions also possible• Eyes always blink. Need to detect changes in
speed of blinking• Need to identify the frames to be used
Solution: Normalize
References• V.manohar, D.B. Goldgof, S.Sarkar, Y. Zhang, "Facial Strain
Pattern as a Soft Forensic Evidence", IEEE Workshop on Applications of Computer Vision (WACV'07),pp 42-42
• Vasant Manohar, Matthew Shreve, Dmitry Goldgof and Sudeep Sarkar, "Finite Element Modeling of Facial Deformation in Videos for Computing Strain Pattern", International Conference on Pattern Recognition, Dec. 2008
• Matthew A. Shreve, Shaun J. Canavan, Yong Zhang, John R. Sullins, and Rupali Patil, "Imaging And Characterization Of Facial Strain In Long Video Sequences",xxxx
• Malcolm Gladwell,” Blink: The Power of Thinking Without Thinking”, Back Bay Books (April 3, 2007)
Thank You!
Sridhar GodavarthyDept. Of Computer Science and Engineering
University of South Floridasgodavar@cse.usf.edu