Research Questions:


How would one implement a system that allows for preference choices to
be made that rank certain attributes of images to enhance certain types of
segmentation? For example, allowing a user in a search application to
specify segmentation by shape first, then color, creating a type of
hierarchy of relevant attributes.


Humans use top-down knowledge to help segment objects in a scene. How could
machine vision systems be endowed with similar behavior? 


Could histograms be used on a more local level to detect changes in regions of
an image (eg, make a grid and take histograms of the grid segments)?


If the segmentation model presented is correct, 
how does the brain address images where there is no clear distinction? For 
example, if a blurred flower is presented to the eye, it appears as a blur 
of colors but can be identified as a flower, where does the brain form 
segmentations.


In determining affinity by texture, there seems to be a problem of 
gradients; a texture changes in scale depending on how close/far the 
point of view is. It would be interesting to see if a clustering method 
could be designed which compared the size of elements of gradient 
textures in the "distance" to those "near" and used this information 
to determine the relative distances of objects in the picture. 



Short Answer Questions:

Describe "agglomerative clustering". 


What is the basic idea behind segmentation strategies?


What is one motivation for image segmentation?


Briefly describe one type of texture labeling method (e.g., Hurst coefficients
or Haralick’s co-concurrency matrices). 


Describe thresholding techniques for image segmentation, including
problems with this approach.



Multiple Choice Questions:

Which of the following is NOT one of the Gestalt principles of 
grouping? 
a. proximity 
b. color 
c. closure 
d. common fate

Which of the following is not a method of texture labeling pixels? 
a. Hurst Coefficents 
b. Haralick's Co-occurrency matrices 
c. Hoover's Formula 
d. Homogeneity


Which is the least reliable way to segment an image?
a) Image Class Templates
b) Histograms
c) Annotation
d) Texture Matiching


Which of the following is not a known classification algorithm? 
a. K-means 
b. PDQ method 
c. Nearest neighbor 
d. Bayesian network


Which of the following is a possible improvement to segmentation by
clustering in image space?
a. Using Hurst coefficients
b. Using luminance thresholding
c. Using quad trees to split regions
d. Using Haralick’s co-occurrency matrix

Yuri Ostrovsky <yo@mit.edu>

1) Research Question:

Do IT neurons exhibit specificity for "clusters" as well as objects?

2) Short Answer Question:

What is the k-mean clustering algorithm?

3) Multiple Choice Question:

Which "illusion" provides the best evidence that "familiar configuration" (tokens tend to be grouped together if, when grouped, they lead to a familiar object) is an important Gestalt principle?

a. The Muller-Lyer illusion (two arrows)
b. The face-vase illusion
c. The dalmation dog illusion
d. The Kanisza triangle illusion (3 pacmen and a triangle)