wk4

Question 1

What is the issue with treating vision as an objective medium

Answer 1

-Vision is inferential and highly dependent on context
-changes in illumination can shift perceived colour of surrounding images

Question 2

What should selected features be invariant with respect to

Answer 2

• illumination
• scale
• rotation
  -affine (stretching of image)
• perspective

Question 3

How do we help with illumination, scale and rotation invariance

Answer 3

illumination: histogram normalization or difference-based shift

Scale: change images to smaller versions via average or mean filter (similar to max pooling)( or use Difference of Gaussian

Rotation: rotate all features in same direction by taking the histogram of all directions and rotate all features to the most dominant direction

Question 4

How could you compress a dynamic video

Answer 4

look for static elements in the video e.g. background which doesn’t change and make the pixels static

Question 5

what is optical flow

Answer 5

visualisation of motion between two images:
-identifies features in t and t+1 which correspond to one another
-plots a vector of movement of each pixel or feature from t to t+1

Question 6

What is a simplistic way to track motion in two images t and t+1

Answer 6

(t+1) - (t) -> if the resulting pixel is greater than some threshold, motion. Else nothing

Question 7

What is the problem with simply subtracting two images to determine motion and what should the solution be

Answer 7

-Two images could be noisy
-noise (being random by nature) will not be correspondent between the two images
-hence subtracting the two images will show differences in noise mistaken for motion
-gaussian filtering will suppress features so is not appropriate
-solution is using connectedness

Question 8

What is 4 and 8 neighbour and connectedness

Answer 8

4 neighbour:
- a feature is only a feature if it shares an adjacent pixel with 4 other features

8 neighbour:
- feature is a feature if it shares a common corner with 8 other features

Connectedness:
-P and Q are connected if a path from P to Q can be joined by a series of pixels such that the path contains 8 adjacent pixels

Question 9

What is the aperture problem

Answer 9

If we only have access to say a slit or a pinhole camera so we see only local features of an image, we cant truly determine the overall motion of an image, directionality nor proper features when looking locally

As a solution we shouldn’t look at local features but instead identify interesting features

Question 10

What interesting feature should we investigate rather than edges

Answer 10

Corners

Question 11

which detector is used for detecting corners

Answer 11

Moravec operator

Question 12

How does Moravec operator work

Answer 12

-Take a filter e.g. 5x5, 7x7 etc.
- Place this filter on some point in the image call it filter A
- Shift the filter in each of the component directions (left, right, up, down, the 4 verticals) call this filter B
- for each B_i where i =[0,8]. Take (A-B)^2
- sum all these differences
- threshold the resulting pixels, the largest ones will be those with the greatest magnitude change in all directions which corresponds to corners

Question 13

what are the 3 principles of motion correspondence

Answer 13

• Distinctiveness: individual points must be distinct from one another
• Similarity: two points should resemble each other if they are the same point affected by motion i.e. point 1 in t and t+1 should look similar
• Consistency: two matches should have moved in analogous ways to other matches

Question 14

outline of the algorithm for motion correpsonence

Answer 14

-find points of interest using Moravec
-pair features of img1 at t and img2 at t+1
- calculate the degree of similarity between each point of interest ,i, in img1 and img2
-calculate the likelihood of each match by calculating similarity weights and converting them to probabilities
-do this for each patch in img1 to each patch in img2, highest probability = most likely match

Question 15

what is the equation for similarity weight from patch i to patch j where patch i is in some image at t and j is in some image at t+1

Answer 15

w_i,j = 1 / 1 + (alpha) S_i,j

where S_i,j is similarity between i and j which is equal to patch_j - patch i

and alpha is a constant

Question 16

Once you have all similarity weights, how do you determine the probability/likelihood of a match

Answer 16

Normalize the weights such that their sum is 1. The resulting weights correspond to the probability of a match

Highest probability is most likely match

Question 17

How do you translate between cartesian and polar coordinates

Answer 17

theta = tan^-1 (y/x)
w = x * cos ( theta ) + y * sin ( theta )

Question 18

How does a hough transform work

Answer 18

• Edge or threshold an image ( can do without it but do it generally)
• Given a point, translate it to its polar coordinates
  -each point has a representation of w and theta, but each point in the Hough space may be represented by different w and theta scaling proportionally
  -For this point in the hough space, find different values w and theta for it and plot lines to generate these points on the hough space w = x cos ( theta ) + y sin ( theta ). Do this for some range of theta e.g. 0-180
• Repeat this for multiple points from the original image, this will generate many lines in the hough space which will intersect at different points
  -points in the hough space with many intersections are likely to represent actual / true lines and edges of the image
  -The same can be done for circles and elipses

Question 19

What are the advantages and disadvantages of the Hough Transform

Answer 19

Hough transform is good at finding lines and locating them in an image

It however does not find where the line starts or ends

Question 20

How could you potentially improve a hough transform

Answer 20

• try suppress local maxima,
• edge thinning
• fill gaps between two lines e.g. if two lines are within some distance, connect them
• use canny / sobel as a basis for the though transform