Taylor Goodhart

1
Taylor Goodhart

• Week 6

2
Background

• Rice University
• Computer Science
• Electrical Engineering
• Mathematics
• 21 / Junior
• Home is Dallas, TX

3
Past Research

• Equithreshold Deformation Spaces of Complex
  Plane Curve Singularities
• Log Canonical Threshhold (LCT)
• Measure of complexity of singularity
• Goal Find set of transformations that preserves
  LCT

Singularity
4
This Week

• MICCAI Challenge
• SPREF for Segmentation
• SPREF as a Feature

5
MICCAI Challenge

• Goal Segment Caudate Nuclei
• SVM
• Features
• Intensity
• Difference between neighbor
• Location
• Libsvm vs. Coresvm
• ALWAYS use coresvm
• Feed results into region growing

6
Results
7
SPREF Background

• SPREF
• Spatiotemporal Regular Flow
• Optical Flow vs. SPREF
• Optical Flow is between two frames
• SPREF considers whole volume
• Spline to approximate flow
• Smoothing effect
• Less memory

8
SPREF Segmentation

• Algorithm
• For each pixel, evaluate SPREF for different
  sized windows
• Calculate standard deviation between curves
• Idea
• At motion boundaries, larger windows will be
  influenced, causing standard deviation to rise

9
Idea
SPREF affected by car
SPREF unaffected by car
10
Examples
11
SPREF Segmentation

• Smoothing effect of SPREF reduces noise
• Next step
• Calculate integral image along each axis
• Calculate derivative of integral image
• Threshhold

12
Example
Intensity of Integral Image will grow faster in
white regions
Threshhold
Integral Image along axis
Derivative of Integral Image
13
SPREF Feature

• Goal SIFT-esque feature
• Rotation Invariance
• Calculate SPREF for whole video
• Rotate frames about SPREF
• Scale Invariant
• For each pixel, calculate SPREF at various scales
• Skew Robust
• For each pixel, calculate SPREF for various
• rectangular shapes

14
Detecting Interest Points

• For each pixel, calculate SPREF at various scales
  in a rectangular window
• Find two scales for which standard deviation is
  maximum

SPREF reflects motion of car AND background.
SPREF reflects motion of car
15
SPREF Feature

• Take local extrema

Better response
16
Benefits

• Highest response will be in areas of constant
  motion
• Comparison 3D SIFT points occur at points of
  abrupt change
• Hybrid? Use both 3D SIFT and SPREF features for
  optimal motion classification

17
Next Week

• MICCAI
• Due by Friday
• Implement
• SPREF Segmentation
• SPREF Feature