Oftentimes, there is a 'trick' that can be used as a way around something. In this case, it may just be temporarily replacing the image with one that has the width of the bottom before checking for intersecting objects, then immediately restoring the original image. If the bottom of the image is not exactly in the center, you may also have to move the actor slightly and then immediately relocate it back after checking for intersecting objects.

For the pixel perfect collision detection class to work, the non-transparent pixels of the images are dealt with. For the 'trick' to work, the transparency is not dealt with at all -- only the extent of the image.

A better way to put it is that the with normal collision detection, only the existence of a pixel is important; with pixel perfect collision detection, the state of the pixels become a factor.

If you mean like taking an exaggeratedly large square transparent image and drawing Mario in the middle of it and then setting that image to a Mario object, then yes. All images are rectangular to begin with, regardless of where the actual color portions within that image are, however, and since normal collision detection uses the extent of the image, not the color portions within it, the hitbox is already rectangular. Most people have problems with the image itself being larger than the color portion within it and want the image size to be reduced without reducing the size of the color portion within it. In other words, they want to strip off all outer rows and columns of pixels that are completely transparent so that the 'hitbox' is at its minimal size.