INTRO
Many people expressed their interest so I've decided to create the first challenge. Hopefully, these bi-weekly contests will help us refine our knowledge and practice new skills. I think it will be a nice thing both for beginners and those people who want to try something new or exchange experience.
THE TASK
In this challenge, we'll face a classification problem. The data set is made of gray-scale pictures of cars. Half of them show cars ( positive example ) and half of them do not ( negative example ). All pictures are in the ".pgm" format. You can use pgmview or other free software to view them.
Positive examples:
http://i.imgur.com/LJoeZuC.png
http://i.imgur.com/Nggvd5U.png
http://i.imgur.com/ywEa7iD.png
Negative examples:
http://i.imgur.com/f5XBQ5O.png
http://i.imgur.com/WUkL1MJ.png
http://i.imgur.com/XHPVXzf.png
DATA SET DOWNLOAD
The data set comes from here ( http://cogcomp.cs.illinois.edu/Data/Car/ ) - I've had to modify it slightly ( thanks alecradford for the information ) I've also included pgm viewer so you can view the images before you do anything.
The data can be downloaded from here:
(https://www.dropbox.com/s/2mg9i6m0gvb8byu/car-detection.rar)
CHALLENGE RULES
Our task is to differentiate whether the picture contains a car or not.
You should train your model on the data from the "training-set" directory and then validate your results on images from the "test-set" folder. Training models on Test Data is forbidden ( for obvious reasons ).
Your task is to achieve as high classification accuracy as possible.
This challenge will accept new contenders up to 24th of April 2014.
You can post your results in comments - along with your method and preferably code so others can replicate and confirm your results. You can edit your post at any time you want, posting your results doesn't prohibit you from updating them when you improve your solution. Moreover, you can post multiple solutions that use different methods!
The person with the best solution wins. After this challenge ends, I will create a post that summarizes the results ( along with a "leader board" - to show each solution accuracy. )
Feel free to get to work - you can use this thread to demonstrate your progress and discuss other solutions. It would be really nice if someone could make a sticky of this.
[link][4 comments]