IPhone library for camera recognition

Question

I hope this falls into the category of "programming issue". I’m all frivolous from Goggling (and read every post here on the topic) on the topic "Computer Vision", but I'm more confused than enlightened.

I have 6 abstract figures printed on a piece of paper, and I would like the camera on the iPhone to identify these figures (from different angles, lightning, etc.).

I used OpenCV some time ago (Java) and I looked at other libraries. The caveat is that they either seem to rely on a jailbroken jailbroken on the iPhone, or they are so experimental and hard to use that I probably end up using day-study libraries to understand that they don't work.

I thought about taking +1000 images of my figures and training the Haar filter. But again, if there is something there that is a little easier to work with, I would really appreciate advice, suggesting people with little experience.

Thank you for any suggestion or advice on what you might have :)

Answer 1

Take a look at the OpenCV SURF feature extraction (they also have a demo version that uses it to detect objects). Surfing functions are characteristic features of the image that are invariant to rotation and scaling. Many algorithms detect objects by extracting such functions from an image, and then use a simple “bag of words” classification (comparing the set of highlighted image functions with the features of your “shapes”). Even without resorting to their spatial alignment, you can have good ones if you have only 6 forms).

Answer 2

While not a library, Chris Greenring explains how iPhone Sudoku Grab recognizes its puzzles in its post here . He seems to recommend OpenCV, not just for jailbroken devices.

Glen Low also talks a little about how Instaviz does its form recognition in interviews for the Mobile Orchard podcast.

Answer 3

I get recognition in my Instaviz iPhone app, and the routines are actually packaged in a library that I call Recog. The only problem is that it is designed to recognize finger or mouse gestures, not image recognition. You pass to the routines a set of points representing the gesture, and it tells you whether it is a square, a circle, etc.

I have not yet decided on the licensing model, but I probably used the minimum license for each location.