Given N SIFT/SURF features extracted from 100 training images describing a specific object, and given M SIFT/SURF features extracted from a new image A, that may or may not contain that object (as well as other objects we don’t care about), how would you determine if image A depicts that object?
The only method I’m aware of is to cluster the training features, and generate a histogram for each training image, and then train a classifier (e.g. SVM) on these histograms. Then you would test for the object in image A by extracting the features, calculate the histogram, and then classify the histogram using the trained classifier.
The main problem with this approach is that it assumes image A only contains the object and nothing else, or it doesn’t contain the object. In other words, if the object were a person, and it were trained on images of that person, it wouldn’t be able to detect that person standing in a crowd, because the resulting histogram would be polluted with the features of all the other people in the crowd.
What are other methods for accomplishing this?
I would recommend to read papers on this topic by the author of SIFT features, David Lowe. See here http://www.cs.ubc.ca/~lowe/pubs.html, search for object recognition.