Had a couple of questions about the LBP Cascade Classifier, with respect to size and aspect ratio
So I've trained up a bunch of these cascade classifiers recently, and they all work pretty well. I had a few questions regarding them, and was hoping someone here could clear a few things up for me.
Firstly, I noticed that the detector has a harder time with objects that take up most of the image. By that, I mean that either the width of the true bounding box is close to the width of the whole image, or the height is. It turns out, a decent chunk of my test data appears this way. In fact, I was able to significantly improve my sensitivity by padding the edges of the image.
While this hack worked out for me, there is a performance downside. the fact that I'm performing detection on the padded regions of the image incurs a noticeable slowdown. Is there a more elegant way of dealing with this scenario?
Second, obviously, the detector is limited to a single width to height ratio during training. The objects that I'm trying to detect, however, aren't. I figure if I want to run the detector with a different aspect ratio, I could resize the image. It works, but this seems like a fairly inefficient and inelegant solution. I was wondering if there was a better way to do this. Perhaps easy some way to stretch the cascade instead of the image?