I am doing doing one of getting started competitions on kaggle, the one which requires you to find coordinates of different objects on images.

I know that the usual method for this is to use sliding window and predict whether some piece of image is that object. But I wonder, what if I just use values of all pixels as features and try to teach a regression algorithm how to predict the coordinates of the object on that image? What kind of regression algorithm would be best?

  • $\begingroup$ Please linke that challenge. Note that getting help is often forbidden in competitions. $\endgroup$ – Raphael Mar 17 '16 at 11:54

It is possible to train a neural network to solve a regression problem. See facial keypoints challenge You should note that training a network for regression is harder then training towards a classification problem. The network is more prone to unnormalised, different deviation values per feature, weight initialisation etc. Try following the tutorial and then you'll have a better grasp on using ConvNets for regression.

Good luck!


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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