I'm not sure, but can I use keypoints of an image for object recognition? I have MATLAB implementation of Harris keypoint extraction. Its output is an array of points, I don't know how can I use these points (the number of points in Harris algorithm vary) for recognizing purpose. There is another method, LoG (Laplace of Gaussian) that produce 120x3 for each image.
Keypoints example:
It's my objects dataset: http://www.vision.ee.ethz.ch/datasets/downloads/Obj_DB.tar.gz , images background is white, as seen in the above image. What I want is training a Neural Network with a train set (some of those pictures) and then test the Neural Network with remained pictures.
If it's not clear I can provide more info.
Did you consider SIFT algorithm?
It compute a unique (Scale and orientation invariant) signature for each "corner" using a variable sized neighborhood. afterwards the signatures can be matched using nearest neighbor (L2 norm).
来源：https://stackoverflow.com/questions/7146584/using-keypoints-to-compute-object-similarity