Multi Categories Object Detection System
Description
Learn and
experiment with the novel object recognition method called “Using image priors
in Maximum Margin Classifiers”. Your goal
is to train classifiers corresponding to different image categories according
to the training algorithm specified by the instructor. Each classifier should
accept (with high probability) objects from its own category and reject objects
from other categories.
For each
category the performance of both linear and kernel classifiers (specified by
the instructor) should be compared. The performance
is measured by counting the false positives and true positives for each category
and for each model.
Data sets
The system should be trained and tested on 101_ObjectCategories data set. This data set contains images of objects belonging to 101 different categories. See Caltech101 web site for details and download.
Main steps
The
project consists of the following steps:
Preprocessing
All images first should be resized to 30x24 pixels. All images have to be normalized to zero mean and a unit length, and transformed to cosine transform (use Matlab dct2.m function).
The Training Step
For each
one of the categories reserve 20 random images from the current category to
create the classifiers.
Since 20
images cannot fully represent the variation in the appearance inside object
class, the recognition results will depend upon the choice of training
examples. To reduce this sensitivity we suggest to sunning several (5) trials
of training/testing and compare the average results. In order to compare models performance it is
important that the randomization will be consistent, i.e. that the same
training examples will be provided to each both models.
Training
code is provided (contact the instructor to obtain
the code)
The CVX_Package must first be installed on your PC (See
appendix A at the cvx_userguide
for instructions).
The Test Step
The test
set should include all the remaining images that were not reserved for the
training. The whole procedure (train + test) should be repeated 5 trials for
validation. On each trial the random choice of training images will be
different.
Results
For each model compute the confusion matrix of a size N on N ( where N is number of tested categories). At the cell k,j report the number of images from class j which were classified as belonging to class k, divided by the size of category j.
Tali Buchnik
Phone or email for setting a
meeting:
Phone: 8301
Email : tbuchnik at cs.haifa.ac.il
Optimization Codes
Contact the instructor
Status