I am looking for a public data-set of images that differ from each other only slightly, so that after applying PCA they can be reconstructed with a small error from very few PCA coefficients. It can be any type of images, the purpose is only to demonstrate an extreme example of PCA.


1 Answer 1


Actually in your case I guess the pure images are not that important. The features that you extract from them are important because if your feature space is constructed base on intensity of images at different picture elements, pixels, then you will need so many coefficients. As an easy solution, use MNIST digits and use shape features to extract features from the images of numbers. You can use plausible number of features and then use PCA for the data that is in the new feature space that you have just constructed. In this case smaller number of coefficients will be needed if the features are fine.

  • $\begingroup$ Thanks but as I said, I want to demonstrate PCA for the actual pixels of images. $\endgroup$
    – elliotp
    Jan 23, 2018 at 12:16
  • $\begingroup$ @elliotp You can use mnist to do so as well but I'm not sure how many coefficients will suffice for you purpose. I guess MNIST suites for your task because most of the numbers are centered, consequently center pixels with each other and marginal pixels with each other will have high correlation which may cause first principal components have great eigenvalues. $\endgroup$ Jan 23, 2018 at 12:26
  • $\begingroup$ @elliotp also as a suggestion, I recommend you to pick sample of images of two different labels and plot the three eigenvalues with the greatest amounts. $\endgroup$ Jan 23, 2018 at 12:48

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.