Transcript Lecture 16

Computational Vision
CSCI 363, Fall 2012
Lecture 16
Stereopsis
1
Random Dot Stereogram
2
Transformed
3
Linear Systems
•Linear functions:
F(x1 + x2) = F(x1) + F(x2)
F(ax) = aF(x)
•Linear systems are nice to work with because you can predict (or
compute) the responses of the system relatively easily.
•For example, if you double the input, the output doubles.
•Fourier Transforms are linear operations. (The Fourier transform
of the sum of two images is the sum of the Fourier transforms of
each image).
•Gabor filters are linear filters.
•Neurons are not linear.
4
Threshold and Saturation
Threshold non-linearity: Neurons do not respond until the input
reaches a minimum level (threshold).
Response
Saturation non-linearity: Neurons have a maximum firing rate.
The response saturates after they reach this maximum.
Threshold
Saturation
Linear response region
Input strength
5
Phase and Half-wave
Rectification
Phase non-linearity: Complex cells are insensitive to the phase
(position) of a grating within the receptive field.
Complex cells do not sum inputs within the receptive field.
Response
Half-wave Rectification: Cortical cells have a low spontaneous firing
rate. There cannot be as large a negative response as a positive
response. The bottom half of the waveform is clipped off.
This can be alleviated with
pairs of matched cells that are
180 deg out of phase with one
another. The difference in
responses acts like a linear
6
time response.
Lateral Inhibition
•There is evidence that a spatial frequency channel is inhibited by
other channels tuned to nearby frequencies. (Also true for
orientation tuning).
•This is accomplished by lateral inhibitory connections within the
cortex, known as lateral inhibition.
•This can cause interesting effects, such as repulsion of perceived
orientation when 2 lines of similar orientation are shown close
together.
•If you adapt 1 spatial frequency, there is an increased sensitivity
at other nearby frequencies.
•Inhibitory interactions can help to make tuning curves narrower.
7
Random Dot Stereogram
8
Binocular Stereo
•The image in each of our two eyes is slightly different.
•Images in the plane of fixation fall on corresponding locations
on the retina.
•Images in front of the plane of fixation are shifted outward on
each retina. They have crossed disparity.
•Images behind the plane of fixation are shifted inward on the
retina. They have uncrossed disparity.
9
Crossed and uncrossed
disparity
1
uncrossed (negative) disparity
plane of fixation
2
crossed (positive) disparity
10
Stereo processing
To determine depth from stereo disparity:
1) Extract the "features" from the left and right images
2) For each feature in the left image, find the corresponding
feature in the right image.
3) Measure the disparity between the two images of the
feature.
4) Use the disparity to compute the 3D location of the feature.
11
The Correspondence problem
•How do you determine which features from one image match
features in the other image? (This problem is known as the
correspondence problem).
•This could be accomplished if each image has well defined shapes
or colors that can be matched.
•Problem: Random dot stereograms.
Left Image
Right Image
Making a stereogram
12
Random Dot Stereogram
13
Problem with Random Dot
Stereograms
•In 1980's Bela Julesz developed the random dot stereogram.
•The stereogram consists of 2 fields of random dots, identical
except for a region in one of the images in which the dots are
shifted by a small amount.
•When one image is viewed by the left eye and the other by the
right eye, the shifted region is seen at a different depth.
•No cues such as color, shape, texture, shading, etc. to use for
matching.
•How do you know which dot from left image matches which dot
from the right image?
14
Using Constraints to Solve the
Problem
To solve the correspondence problem, we need to make some
assumptions (constraints) about how the matching is
accomplished.
Constraints used by many computer vision stereo algorithms:
1) Uniqueness: Each point has at most one match in the other
image.
2) Similarity: Each feature matches a similar feature in the other
image (i.e. you cannot match a white dot with a black dot).
3) Continuity: Disparity tends to vary slowly across a surface.
(Note: this is violated at depth edges).
4) Epipolar constraint: Given a point in the image of one eye,
the matching point in the image for the other eye must lie
15
along a single line.
The epipolar constraint
Feature in
left image
Possible matches
in right image
16
It matters where you look
•If the observer is fixating a point along a horizontal plane through
the middle of the eyes, the possible positions of a matching point in
the other image lie along a horizontal line.
•If the observer is looking upward or downward, the line will be
tilted.
•Most stereo algorithms for machine vision assume the epipolar
lines are horizontal.
•For biological systems, the stereo computation must take into
account where the eyes are looking (e.g. upward or downward).
17