Typically, all input images are scaled to a fixed widthxheight, and the non-used pixels around them are cropped away so that new images look like the training set. This resolves issues with scale and transformations somewhat, although it yields difficulties of its own such as improper stretching or the disappearance of thin lines.

Very common problem. You'll be glad to know there is a simple solution. :)

A common approach to this is to look through the image at different scales, and feed the scaled data in to your detection algorithm.


Lets pretend our algoritm examines a 10x10 block of image data. The picture we're looking at is 100x100.

First, break the image up in to 100 10x10 blocks, and check each block.

Then we break the image up in to 20x20 blocks, scale the 20x20 blocks so they are 10x10, and check again.

Then 30x30, 40x40, 50x50, and keep going until we get to the true size of the source image (100x100), each time scaling the 'blocks' down to the size our algorithm wants.


This is a good approach if your object detection routine spits out a 'yes / no' type of answer.

You may run in to the problem of detecting the same object more than once, but you can use some simple logic to post-process the results and identify the true location.

I've actually been working (slowly but surely) on an object detector of my own, using a slightly different approach. Instead of having a boolean type output, I'm attempting to get the algorithm to tag pixels in the source image that it believes belong to a face. I don't actually want it to tell me _where_ a face exists, just that a group of pixels probably belongs to a face. There is a subtle difference. Anywho, I'm hoping that I'll be able to overcome the problem of scale this way.

I'll then use a second program to segement the original image in to separate faces by processing the output of the first algorithm.