Project 3: Object categorization with using "The weak shape model".
Description
Learn and improve a novel method for comparing images
of objects. The method proposes a new distance between images that incorporates
local appearance of object parts and some information about the geometrical
relation between these parts. The proposed distance is incorporated in SVM as a
kernel. The basic code implementing this
method will be provided. The project is mostly experimental and it consists of
the following tasks:
1) Tuning
free parameters of the method, specifically
- the size
of " code books ".
- the
number of "components".
2) Conduct
recognition experiments on Bar-Hillel data sets:
1) Animal set: "dogs against easy animals
" and "dogs against hard animals" .
2)
Furniture set: chairs against other furniture.
Results
The detection results should
be presented as a confusion matrix for each size of "code book " and
size of "components"
and for each of the data sets mentioned above.
For example:
Size of code book 50,
number of components 12
|
Easy Animals |
Dogs |
|
|
How many of the easy animals are classified as Dogs |
How many of the Dogs are classified as Dogs |
Dogs |
|
How many of the easy animals are classified as easy animals |
How many of the Dogs are classified as easy animals |
Easy Animals |
Obviously the best results will produce diagonal confusion matrix.
Data sets
The system should be trained
and tested on Bar-Hillel data sets:
1) dogs vs. easy animals
2) dogs vs. hard animals
3) Chairs vs. furniture background.
The data will be provided by
the instructor.
Mr. Elran Morash
"
Consulting hours:
Monday
10:00-12:00 a.m
Wednesday 16:00- 18:00 p.m
Phone: 8301
Email: elranmorash at nana.co.il
Implementation software:
Contact the instructor.
Useful Links
Class lecture on object
detection
"Distinctive image features from scale-invariant keypoints" by David Lowe
Status