Car Recognition

Description

A tool which will analyse submitted images, and attempt to classify them through image recognition by manufacturer. It will be built on top of the OpenCV library.

Classifiers:

• classifier by manufacturer
  • does it need to be by model type?
    • is generic Porsche (for example) classifier sufficient?
    • or do need 996, 997, Cayman, Boxster etc
  • separate classifiers for angle types (front, side, rear etc)
• HAAR or LBP?
  • LBP potentially much quicker to train?
  • Which has greater accuracy for given training time?
    • (LBP could be trained with more images in same time?)
• Accuracy?
  • how many training images are sufficient to get >70% accuracy
  • curve of accuracy… graph on time to train vs accuracy over different sized sets of prospects.
• online training of classifiers?

Application:

• take input files (image), add to queue.
  • Should it take input over HTTP or direct from disk?
  • initially HTTP I think, because not done it before.
• have classifier workers pull images off the queue, attempt each form of classifier in turn (most populate types first)
• worker output is:
  • If single classifier matched, HTTP response with classifier that matched.
  • if multiple classifiers match, then review confidence for each classifier
    • respond with both along with confidence for each.
  • if no classifiers match, then respond saying no match, flag for manual intervention.

Background Reading:

• http://www.vision.ee.ethz.ch/~hegrabne/papers/Grabner2008TrainingSequentialOn-lineBoosting.pdf
• http://en.wikipedia.org/wiki/Online_machine_learning
• http://web.mit.edu/seyda/www/Papers/GHC06_ACMSRC_abstract.pdf
• http://lmb.informatik.uni-freiburg.de/papers/download/fe_za_bu_GFKL07.pdf
• http://www.svms.org/training/ChLi.pdf
• http://www.csie.ntu.edu.tw/~cjlin/papers/newsvr.pdf
• http://dsp.stackexchange.com/questions/3149/online-boosting-training-on-the-fly-for-face-detection
• http://stackoverflow.com/questions/8791178/haar-cascades-vs-lbp-cascades-in-face-detection
• http://opencv-users.1802565.n2.nabble.com/LBP-vs-haar-cascades-td7312041.html
• http://hal.inria.fr/docs/00/62/43/60/PDF/Finalversion_Haar_like_and_LBP_based_features.pdf