About The Exam
The format of the exam is a mock peer-reviewed conference. You will develop an algorithm, prepare a professional paper, submit an anonymized version to the EasyChair conference system, and peer-review the work from other groups. There are three deliverables:
Phase 1: Develop your algorithm and write your writeup
Develop your algorithm and submit your results to the Kaggle leaderboard.
Deadline: Monday, May 15, 11:59 PM EST.
Write your paper submission and submit to Dropbox (link coming soon!).
Deadline: Monday, May 15, 11:59 PM EST
. There should be only one submission from each team.
Phase 2: Peer-review other groups' work!
Write reviews and submit them to CMS.
Deadline: Tuesday, May 16, 11:59 PM EST
. (Will open after Phase 1 ends)
Everybody
on each team must complete peer reviews individually. (They are not "per-team")
The Challenge: Build a large-scale image search engine!
You and your team of
two Cornell Tech students
are surely on the path to fame and fortune! You have been recruited by Google to disrupt Google Image Search by building a better search engine using novel statistical learning techniques.
The specifications are simple: We need a way to
search for relevant images
given a natural language query. For instance, if a user types "dog jumping to catch frisbee," your system will
rank-order the most relevant images
from a large database.
Here are some details:
During training,
you have a dataset of 10,000 samples. Each sample has the following data available for learning:
A 224x224 JPG image.
A list of tags indicating objects appeared in the image.
Feature vectors extracted using
[ResNet](https://arxiv.org/abs/1512.03385)
, a state-of-the-art Deep-learned CNN (You don't have to train or run ResNet -- we are providing the features for you). See
[here](http://ethereon.github.io/netscope/#/gist/b21e2aae116dc1ac7b50)
for the illustration of the ResNet-101 architecture. The features are extracted from
pool5
and
fc1000
layer.
A five-sentence description, used to train your search engine.
During testing,
your system matches a
single five-sentence description
against a pool of 2,000
candidate samples
from the test set. Each sample has:
A 224x224 JPEG image.
A list of tags for that image.
ResNet feature vectors for that image.
Output
: For each description, your system must
rank-score
each testing image with the likelihood of that image matches the given sentence. Your system then returns the name of the top 20 relevant images, delimited by space. See "sample_submission.csv" on the data page for more details on the output format.
Evaluation metric
: There are 2,000 descriptions, and for each description, you must compare against the entire 2,000-image test set. That is, rank-order test images for each test description. We will use
MAP@20
as the evaluation metric. If the corresponding image of a description is among your algorithm's 20 highest scoring images, this metric gives you a certain score based on the ranking of the corresponding image. Please refer to the evaluation page for more details.
Use all of your skills, tools, and experience. It is OK to use libraries like numpy, scikit-learn, pandas, etc., as long as you cite them. Use cross-validation on training set to debug your algorithm. Submit your results to the Kaggle leaderboard and send your complete writeup to EasyChair. The data you use --- and the way you use the data --- is completely up to you.
The best teams (of
two Cornell Tech students
, recall) might use visualization techniques for debugging (e.g., show top images retrieved by your algorithm and see whether they make sense or not), preprocessing, a nice way to compare tags and descriptions, leveraging visual features and combining them with tags and descriptions, supervised and/or unsupervised learning to best understand how to best take advantage of each data source available to them.
The best reports will be professionally written. We suggest including an
Abstract,
followed by an
Introduction
section including motivations of your method and a detailed description of each step in your pipeline / framework you used for this task. Then, include a detailed
Experiment
section and a
Results
section. If time permits, a
Background / Related Work
section can help to point out the relevant work in the image retrieval literature.
The best reports use a professional style adopted by a major academic conference. NIPS is a good choice. Teams can download a template from Section 4 Paper Format of the
[NIPS Author Guidelines](https://nips.cc/Conferences/2015/PaperInformation/AuthorSubmissionInstructions)
. We recommend you to use the LaTeX template, but this is not a requirement. If you're more familiar with Word, use Word.
The best peer reviews will be detailed and thorough, pointing out potential areas of improvement as well as complimenting their peers' strengths.
The best teams understand how to divide the precious time and energy among all members of the group. They might split into roles, preventing programming students from messing up the writing and ensuring writing gurus don't screw up the code. Close contact is always maintained within top-performing teams. When hardship happens within the best teams, blame is not assigned---it is overcome.
May the best team win!
Rules
You are strongly encouraged to work in a group of
two
students.
Groups may not collaborate with other groups. You cannot cite other teams' work. This includes sharing data, intermediate and final results, writing, or discussing your detailed method with other groups. You
may not
view reports from other teams until
after
the peer-review phase begins!
How to anonymize your paper
The "Peer Review" phase will be double-blind. This means authors will not know who the reviewers are, and the reviewers should not know who the authors are. This protects both sides.
To maintain anonymity,
don't
include team member names in the actual paper that you submit to Dropbox!
However, we need to know who you are. So please use the following file name:
TeamNameOnKaggle_Author1lastname_Author2lastname_report.pdf
This will be hidden from your reviewers.
![](http://i.imgur.com/F255rFT.png)
Peer Review Tips
See the following resources for tips on how to write a professional review:
Matthew Might’s notes on
[how to peer review](http://matt.might.net/articles/how-to-peer-review/)
and
[Peer Fortress: The Scientific Battlefield](http://matt.might.net/articles/peer-fortress/)
(the last one provides examples NOT to follow)
[CVPR 2016 Reviewer’s Guidelines](http://www.pamitc.org/cvpr16/reviewer_guidelines.php)
.
[So You're a Program Committee Member Now: On Excellence in Reviews and Meta-Reviews and Championing Submitted Work That Has Merit](http://mobilehci.acm.org/2015/download/ExcellenceInReviewsforHCICommunity.pdf)
Score Breakdown
30% on Kaggle performance (score on the private test set).
30% on report: is the proposed method clearly described? is it professionally written? does it include enough detail for a professional (ie. skilled graduate student) to re-implement the results? A good report should at least contain a brief introduction of what you did, a detailed description of your method, an experimental evaluation part that shows the experimental results and analysis of your results.
20% on method: does it make sense? is it suited to the task? does it make use of the data in appropriate ways?
10% on peer reviews from other classmates.
10% on the quality of peer review: is it well-written and thoughtful? does it provide insight to the authors?