Charisse Massay <charisse@mit.edu>
1. Open Research Question: I can't think of an open research question, I
would have to think of a different method of recognition and test that.
I'm not sure of how to move past the descriptions offered for recognition.
2. Short Answer Question: What are the three major models of recognition
and what is the basic principle of each?
3. What is a 'gaussian'?
a. a penalty for deviations
b. center peaks in Radial Beta Functions
c. level of correspondence between the image, model and transformation
d. unknown
---------------------------------------------------------------------
Matt Cain <mcain@mit.edu>
Future research:
The obvious research to come from this would be to implement it in a computational system and try to identify objects with it. Start off with isolated objects and then go to more cluttered scenes.
Short answer:
What formula is used to give the 'fit' of the alignment approach?
Multiple Choice:
In the HyperBF Network Model, what is the function trying to maximize:
a. deviations
b. derivatives
c. smoothness
d. speed
---------------------------------------------------------------------
Roland Fleming <roland@psyche.mit.edu>
-- Multiple Choice:
Which of the following is not true ?
(A) The alignment approach to recognition uses 3D object representations.
(B) Feedforward models of recognition are always more effective than
recurrent networks.
(C) The HyperBF scheme uses radial basis functions to solve the
regularization problem.
(D) By linear combination of views it is possible to capture 3D transforms
even though the input and stored models are 2D images.
-- Short answer question:
Briefly explain the main assumptions behind the HyperBF algorithm.
-- Research issue:
The models of recognition discussed in class all deal solely with
within-category identity recognition, such as working out which, from a set
of bananas, a given novel view corresponds to. How can these models be
adapted to classify entirely new objects, that is, objects not present in
the set of stored representations? How can the models be improved so that
when an object is recognized as new, it is added to the set of stored
representations? And, most difficult of all, how can these systems extract
the basic categories of objects from a series of views?
---------------------------------------------------------------------
"Jennifer C. Shieh" <jcshieh@mit.edu>
1. Is there an implementation or a variation of the HyperBF network model
that could take 2D images as input and output a 3D model or a
representation of 3D information?
2. Name and briefly describe three models/theories of recognition.
3. Which of the following theories of recognition is best supported by
biophysical evidence?
a. Alignment approach
b. Linear combination of views
c. HyperBF network model
d. Object decomposition methods
---------------------------------------------------------------------
Richard Russell <rrussell@mit.edu>
1) open research question/project idea:
It would be useful to differentiate the abilities and disabilities of these
models from one another in order to look for empirical evidence in support
of one or another.
2) Short answer:
What assumptions are common to the alignment, linear combination of views,
and hyper basis function networks models of object recognition?
3) Multiple choice:
Which of the following are not necessary steps of the alignment approach?
A) linearly transform the image to match model
B) establish correspondences between image and model feature points
C) transform the model to the image on basis of correspondences
D) decide whether image and transformed model match
---------------------------------------------------------------------
Kennet Belenky <kbelenky@mit.edu>
Open question:
There are various feature extraction algorhythms (sp?) that have been
proven to work in the domain of face recognition. (AdaBoost, PCA, HMAX).
Investigate if there is any relation between the features extracted by
these algorhythms and the features detected by animal/human visual systems.
(and if you conclusively find something, you'll be in Stockholm tomorrow
accepting the Nobel prize).
Short answer:
Most statistical feature extraction algorhythms are not very good at
position and rotation invariance. Discuss possible ways to augment feature
extraction systems to create a more robust system. What are the
implications of implementing this system with neurons (parallel processing
structure) or a computer (serial processing)?
What are the benefits of a support vector machine over a simple nearest
neighbor(s) system:
a) Improved calculation time.
b) improved accuracy
c) improved training time
d) All of the above.
---------------------------------------------------------------------
Jodi Davenport <jodi@MIT.EDU>
1. Open question:
Which aspects of the three models discussed are most closely related to
the way the brain operates.
2. Short Answer:
Why do corresponding points need to be found in the linear combination of
views model?
3. Multiple choice
Which of the following is not a model of recognition?
a. The transfiguritron approach.
b. The alignment approach.
c. Linear Combination of views.
d. Hyper BF network model.
---------------------------------------------------------------------
Yuri Ostrovsky <yo@mit.edu>
1) Research Question
Does explicit knowledge about the 3D structure of a face or object aid humans in their recognition of those faces or objects?
2) Short Answer Question
What is a radial basis function?
3) Multiple Choice Question
Which of the following is NOT one of the image transforms in the ALIGNMENT approach for object recognition proposed by Ullman and Huttenbocher (1990)?
a. Align the center of mass with the center of view
b. Orient the image
c. Scale the image
d. Sharpen the image