Project 1 - Image prediction
Due date: Oct 8 2018, 1:00am
This project focuses on image level prediction using deep (convolutional) networks. In this project you will train a network to predict a single label from an image. You have two choices for datasets: Age prediction from yearbook photos or geolocation prediction from streetview images.
Face to age
This dataset contains frontal-facing American high school year-book photos with labels to indicate the years those photos were taken. Your task is to train a deep network to predict the year a novel image was taken.
Dataset credit Shiry Ginosar
This dataset contains images from France street view with geolocation label (latitude and longitude). You task is to predict the geolocation of novel street view images.
Dataset credit Carl Doersch
Starter code and data
Check the starter code out here.
For both project, we provide you with a training data set and a validation data set (you can use it as your test set). However, we will keep a hidden test set to make sure your algorithm is not overfitting too much.
We also provide you with a python script named
run.py which defines the interface we will use to evaluate your code on the test set. You can modify
run.py to run your own model (including loading your parameters). See the comment in script to know how we will test your code. Make sure your
run.py works with the evaluation code we provide without any modifications to the evaluation code. Otherwise we will not be able to evaluate your code and you might not get the full credit.
The starter code also contains scripts to load the images and labels for your convenience.
- A writeup. Preferably in latex using the CVPR 2018 stylesheet. It should contain:
- An intro and abstract outlining the method you chose and why it’s better than alternatives.
- A technical section, with a detailed overview of your method and any extensions.
- A thorough evaluation of any aspects of your method. Try to justify every choice you made either with a sound argument plus citation or with an experiment.
- The code and parameters to run your model. This code should be written in python. Contact the TA if not.
How to submit
Please submit a pdf of your writeup, your results on the test set and your code in a single compressed file named as
[UTEID]_[You name]_[yearbook/geo].[zip/rar/tar.gz] (e.g.
geo mean which project you choose. If you decide to try both please submit two files (with the same writeup file).
The compressed files should include:
run.pyyou can add anything, just to make sure the TA can get predicted label by calling predict(image_path) in class Predictor.
- A report named
report.pdf, as detailed above.
- Any dependencies you used that the
run.py(e.g. trained model file, other auxiliary python files)
- A text file with the results on the test set.
- Put Model and generated test label file in the respective folder.
- Add readme.md file in your submission with proper description of each files.
- Project should be run from the grade.py file as shown in the evaluation step and should be able to generate the test label file. Make sure while generating test label file, do not shuffle the data.
- There should not be any mismatch between the test label file and the label predict by the model. Otherwise, your entry will be rejected.
- Keep the submission under 2GB. Otherwise, email the TA.
- Do not include any dataset (image) files in your submission
- Please make sure that in average each prediction takes less than 500ms on a Nvidia GTX 1080 or equivalent GPU. Try to not load the model over and over again (load it just once). Do not train your model on the TAs machine.
- The program will be terminated if it exceeds the timing limits
- Test your program before submission to make sure the run.py works. Also make sure that there are no absolute path names or relative path names to files that are not submitted.
Each team has 7 min to present. Not every team member needs to speak, if you don’t want. A good presentation should spend:
- 1 min or less on explaining the method
- if common with other teams
- Focus on what you think sets your team apart from the competition
- Did you face some interesting challenges, that you could or could not master? (installing tensorflow does not count!)
- Did you do anything fun in addition to just the basic task?
- If your method works incredibly well, what did you do?
- (optional) the presentation can be interactive
- ask how your class-mates solved certain aspects
- how it compares to your approach
This project counts 25% towards your final grade. The 25 points are split up as follows:
- 7pt basics
- 4pt writeup
- Technical section
- Basic evaluation
- 3pt twice chance performance
- Evaluated by absolute distance between prediction and ground truth.
- Chance is defined as the best (deterministic or stochastic) predictor that completely ignores the input image.
- 4pt writeup
- 8pt extensions (any of these will give you 8pt, if executed successfully). Extension ideas include, but are not limited to
- Visualizing the important regions in the image used by the network
- Comparison of different losses (more than two), such as regression or classification
- Benchmarking different model architectures
- Training your own model architecture from scratch. You need to justify (theoretically or experimentally) why it might work better than standard architectures.
- Your own idea. Double check with the instructor or TA if you’re uncertain if it’s worth 8pt.
- 5pt running both geolocation and face2year (optional)
- 3pt best two submissions on test set (optional)
- Either geolocation or face2year
- 10pt presentation
If you get more than 25pt we truncate the project 1 score at 25% (including presentation).
Late submission policy
The presentation cannot be late, you will loose 10pt if you do not have a presentation ready. You will lose 2pt for every day the submission is late. You will also not be eligible for the best submission score.