You are on page 1of 2

Final project:

- First of all, you need to determine a method that has been put forward in the last 5 years
(2018 and later) (method or network model must be "proposed" in the study, it is not valid if
the known or previously proposed method has been applied in another area). For this, I
recommend you to review publications published in high-level conferences (such as ICPR,
NeurIPS, CVPR, ICML, KDD) and high-level journals (such as IEEE TPAMI, Pattern Recognition,
JMLR, Neurocomputing). It is useful that the method has a level of complexity that you can
understand (I am undecided about the oral interview, maybe, but a basic understanding is
required for code adaptation). In addition, sharing the ready-made codes of the method (via
github or sites such as paperswithcode.com) will be easier for you (the aim is not to code the
method or model from scratch, but to gain the ability to adapt it to a different area with all the
necessary adjustments / pre-processes by understanding it well. ). You are free in the
programming language/libraries.
- In order to clarify the subject, you must separately approve the method in the publication
you have chosen and the data set you will use. I will process the subject and data sets that I
will approve in the file that I have shared with you before. Deadline for approval: Tuesday,
November 15, 2022. For students who have not determined a subject and data set by this date,
an additional period of 1 week is given by deducting 10 points. Those who do not choose a
subject within that period are deemed not to have taken the final exam.
- For datasets I'm looking for the requirement that is not small in scale (In case of Tabular
dataset, it should not be less than 100000 samples and 20 features. If Image dataset, it should
not be less than 10000 images (of course, if the image sizes are too large, the number of
images may decrease)) and that the method you choose has not been applied on a very similar
dataset. Therefore, within the scope of the project, I will want to apply an up-to-date method
that can be considered state-of-the-art to a different field (dataset). Example for a different
field: Although it is not accepted to apply a method applied to the aaa set in face recognition to
the bbb set, you can apply it to a data set (if not applied) in palm print recognition with
different dynamics.
- One issue that I care about in the evaluation: how much the students examine the effects of
the hyperparameters of the method and whether they apply pre-processes etc. for compliance
if necessary. It will not directly affect the score if the results are very good, but if there are good
results (and if you want), we can publish them after the semester is over.
- You can share with data@cs.ktu.edu.tr or murataykut@ceng.ktu.edu.tr for submission at the
end of the semester (also valid for midterm homework).
- Your working codes and min. I want you to write a Turkish paper consisting of 3 pages (no
maximum limit). The titles I want here are not in full paper format: abstract, methodology, data
set, experimental results and interpretation titles are appropriate. There is no need for an
Introduction section with a literature review. You can use the format of the SIU Congress as the
format (doc / tex file).
- If the project is the same as a project presented/to be presented in another course, a "zero"
score will be awarded directly (I will be in contact with other professors).
- I will not accept topics directly related to computer vision such as Object detection /
segmentation / tracking as a project topic within the scope of this course. Similarly, the topics
of smart optimization / evolutionary algorithms are smart opt. / not suitable because it is
associated with soft computing courses. You can perform recognition (classification) / clustering
on images.
- For data set approval, I want you to write summary information besides the link: How many
samples, how many classes it contains, how the samples were taken (for example, how many
sessions), information such as the resolution of the image.
- I will not check whether the method and the data set will be compatible. Therefore, it may be
better to select the dataset after learning a little about the method.
- It is your responsibility to check that the method you will use for the final project has not been
implemented in the area associated with the dataset (for example, PolyU Palmprint DB-II) (for
example, palmprint recognition) (it is not enough that it is not used only in that dataset, but if it
is used in other similar datasets in the relevant area, it is not suitable for the project). I will not
check when accepting project topics / datasets but if I see it used before project evaluation you
will get 0 points straight away.
- Let me make a small caveat for the dataset: Yes, I do not want small datasets, but you should
not choose datasets that require too much of your computers' processing power / GPU support
of the cloud system you are using, or that require an unacceptable amount of test time.
- Let me list some links below that can help you find datasets (you don't have to stick to them,
you can get datasets from other sources related to your field of study):
-- https://archive.ics.uci.edu/ml/index.php -> most of the datasets on this site are small and
invalid, but there are also large ones (I am sharing because it is a basic resource for ML)
-- https://guides.library.cmu.edu/machine-learning/datasets
-- https://www.v7labs.com/blog/best-free-datasets-for-machine-learning
-- https://www.datasciencecentral.com/profiles/blogs/big-data-sets-available-for-free
-- https://homepages.inf.ed.ac.uk/rbf/CVonline/Imagedbase.htm
-- http://www.cvpapers.com/datasets.html
-- https://en.wikipedia.org/wiki/List_of_datasets_for_machine-learning_research

You might also like