ECE 5554 / ECE 4554: Computer Vision Fall 2017
Overview
The final project is a chance to further explore a topic of interest. Groups
of up to four are highly encouraged. More is expected of larger groups. Projects will include
a project report webpage and a poster presentation. Various types of projects are possible.
You could implement a paper that you find interesting, something discussed in class, a
significant extension of one of the course projects, or something entirely of your own design.
The work does not have to be of publishable originality. However, you are encouraged to
submit the revised versions of projects to top computer vision conferences.
Research project: Perform a project in a topic of your choice. Formulate a goal, devise
an approach, and evaluate. When proposing, indicate what dataset you will use for
evaluation. For example, you could base your project on an existing paper and try to
improve the accuracy or speed with some modification. You could also apply existing
algorithms to your own field (e.g., robotics).
Review and implement a paper: Choose a paper or set of papers and write a scholarly
review. Then, implement and evaluate the algorithm. If done in a group, more than one
paper should be implemented and compared. Reviews should be written independently
for each person, but the group can collaborate on implementation and evaluation.
Note that you cannot directly use your previous research project/papers as your final project.
Your final project can build upon your previous work but you must clearly specify the what are the new efforts carried out in this semester.
Violation will result in 0 credits.
Grading
Proposal (10% of project grade):
A 1-page description that describes the following:
Problem statement: Clearly state the goal of your project. When someone uses your system, what is the expected input to the system, and what is the desired output?
Approach: Describe the technical approach you plan to employ.
Experiments and results: Describe the experimental setup you will follow, which datasets you will use, which existing code you will exploit, what you will implement yourself, and what you would define as a success for the project. If you plan on collecting your own data, describe what data collection protocol you will follow. Provide a list of experiments you will perform. Describe what you expect the experiments to reveal, or what is uncertain about the potential outcomes.
Project Report Webpage (50% of project grade):
We expect you to have completed your project by this time.
The only thing we expect between this and the final poster presentation is incorporating feedback from the instructor and/or TA.
The report should be in the form of a webpage. It should be a nice, professional looking, visual, self-contained webpage describing your project.
We will link to all project pages from the class webpage.
Any student in the class should be able to clearly understand your report such that he/she can implement it.
Below is an outline of what we expect the webpage to cover:
Abstract: One or two sentences on the motivation behind the problem you are solving.
One or two sentences describing the approach you took. One or two sentences on the main result you obtained.
Teaser figure: A figure that conveys the main idea behind the project or the main application being addressed.
Introduction: Motivation behind the problem you are solving, what applications it has, any brief background on the particular domain you are working in (if not regular RBG photographs), etc. If you are using a new way to solve an existing problem, briefly mention and describe the existing approaches and tell us how your approach is new.
Approach: Describe very clearly and systematically your approach to solve the problem.
Tell us exactly what existing implementations you used to build your system.
Tell us what obstacles you faced and how you addressed them.
Justify any design choices or judgment calls you made in your approach.
Experiments and results: Provide details about the experimental set up
(number of images/videos, number of datasets you experimented with, train/test split if you used machine learning algorithms, etc.).
Describe the evaluation metrics you used to evaluate how well your approach is working. Include clear figures and tables, as well as illustrative qualitative examples if appropriate. Be sure to include obvious baselines to see if your approach is doing better than a naive approach (e.g. for classification accuracy, how well would a classifier do that made random decisions?). Also discuss any parameters of your algorithms, and tell us how you set the values of those parameters. You can also show us how the performance varies as you change those parameter values. Be sure to discuss any trends you see in your results, and explain why these trends make sense. Are the results as expected? Why?
Qualitative results: Show several visual examples of inputs/outputs of your system (success cases and failures) that help us better understand your approach.
Conclusion and future work: Conclusion would likely make the same points as the abstract.
Discuss any future ideas you have to make your approach better.
References: List out all the references you have used for your work.
Examples:
See this for a webpage template.
See this for an example of a nice, professional looking page.
See this for an example of how to lay out the various details of your project. You may need to provide more details than this, beause you will not be submitting an associated paper to accompany the webpage. So the page should be self-contained.
Poster Presentation (40% of project grade):
A poster presentation that describes the same points as the proposal and the project report webpage.
A sample poster template can be downloaded from here.
Describe the results you have obtained.
This may involve some additional experiments, etc. based on the instructor's and/or TA's feedback on your report webpage.
Live demos when applicable are highly encouraged. Prepare a 1 minute schpeel that presents the main summary of your work. The instructor and TA will ask further questions as necessary.
Some final project ideas
Try something new or interesting: apply vision to a mobile robot, make a data-driven interface
for image editing, organize home photos, do visual search in a home photo collection,
reconstruct a 3D scene from multiple images, estimate material properties, etc.
Compare two or more approaches: Implement two approaches (e.g., for object recognition)
and try to understand when one works better than the other. As part of this, you could try to
create a better benchmark dataset (maybe a smaller prototype).
Some specific ideas if you need one (though I prefer if you come up with your own idea):
Face tracking in videos: Tracking facial landmark in videos
Player tracking: Try to track players and the ball in a sports video.
Action recognition: Try to detect when somebody is performing a particular action in a video, e.g., hug, shaking hand, kiss.
Pedestrian detection: Build a detector for standing/walking people
Multiview reconstruction: Build a system to reconstruct an object or a scene from multiple images
Rubik cube recognition: Use a webcam to detect all small square and cube orientation.
Similar category differentiation: Make a classifier that can tell the difference between
dogscats or bicyclesmotorbike, different kinds of birds, etc.
Material detection: Try to classify materials on natural objects in images.
Recognizing art/photo styles: Build a classify to recognize different styles.
Gender/age classification: Given a face, try to predict the age and gender.
Fake or Real: Try to predict whether an input image is natural or was generated by a computer.
Im2Calorie: Try to estimate the type of food and its calorie.
Text detection: Spoting texts in the street.
Shadow detection: Try to find and remove cast shadows in outdoor images
Photo organization: Build a system that can organize your photos by the people in them.
Evaluation of object detection: Run a state-of-the-art detector on a PASCAL VOC dataset or MS COCO
and study (quantitatively and qualitatively) which factors make detection difficult.
Credits
Project description by Devi Parikh and Derek Hoiem
|