Hi
I need to implement a car make and model recognition system. I assume to have my unknown car cut out from the scene quite well (and normalized), but not perfectly (so simple MatchTemplate won't work well, unless I generate many slightly distorted images, but that would take ages). My ideas:
- Build a database of car (fronts) keypoints and compare keypoints of the new image with FLANN - probably slow.
- Use the eigenfaces method to model car fronts.
- Build a database of downsampled and normalized car fronts and use pixel values directly to build some (tree-based?) classifier - probably would need a very large training set (just guessing).
- As in 3, but DFT or DCT the images and select a number of most significant values..
Do you have any comments or suggestions?
Thank you
Witek