Open source object recognition algorithms

I've been doing a lot of search on the right algorithms to use for my object recognition program. But everything I've found had some major flaws.

My program is supposed to learn new objects as it encounters them, online. When it encounters a new object, a box is bounded over it and the object is learned(OpenTLD does this perfectly). Reiterate this 1000s of times for different objects, the program should be capable of recognizing 1000s classes and instances of objects(Haar-like feature cascade can do this, OpenTLD fails). The algorithm must be scale and orientation invariant(Haar fails).

Everything I've found could only fulfill some of the above criteria, while failing the rest. Surprisingly I've yet come across anything that could m开发者_JS百科eet all the criteria. I only mention Haar and OpenTLD above as they are the ones that are the closest to what I need. Other algorithms, like SIFT, SURF are even farther from what I need.

So my question is, are there any existing source code out there that does what I need? Or is this something I would have better luck just modifying existing source code with?

TLD and Kalal's work is excellent for tracking. However, recognition is a very different problem.

I don't understand why you dismiss SIFT or SURF for recognition. I would definitely look in that direction. Histograms of Oriented Gradients (HoG) (Wikipedia) is a related family of algorithms that is considered state-of-the-art. It is used in one form or another in a lot of ICIP'11 papers for example.

What you are asking is still pretty much an unsolved problem in computer vision. You won't find much ready-to-use code. Research code (e.g. Matlab), for non-commercial uses, yes, but you may have to modify/optimize it a lot for real-world use. Some companies like Pongr, Kooaba, IQEngines offer online APIs for object recognition.