Object Matching

An object matching method based on Lowe, D.G. Object recognition from local scale-invariant features

Description

Object matching method based on Lowe, D.G. Object recognition from local scale-invariant features. Marks the contour of the target in a test image based on 1 target image. Uses SURF points instead of SIFT points.
 
Need to download SURFmex library first.
See http://computervisionblog.wordpress.com/2012/01/15/object-matching-method-made-in-the-20th-century/ for detail.

• Required Products: MATLAB
• MATLAB release: MATLAB 7.6 (R2008a)

[Tolga Birdal]

Good work. Could you share the matching algorithm you used in this example? Is it published?

Thanks,

[Li yang Ku]

Thanks, I added some brief description of the algorithm inside my blog post
http://computervisionblog.wordpress.com/2012/01/15/object-matching-method-made-in-the-20th-century/
The algorithm is published in the paper I mentioned inside the blog post.

[Panimalar]

Sir,
    I want matlab code to implement these SIFT algorithm in video sequence.
My mail ID [e-mail]

[Li yang Ku]

I am not sure if this algorithm is fast enough for real time, but you can always sub-sample your video and use the same code in a for loop.

[Panimalar]

Thank you.
Sir, can I get the full published paper on above object matching, because I cannot find it on the blog you mentioned. Can u please mail the same to me.My ID [e-mail]

[Saleh]

Hi,
when I apply your algorithm on my images, it is unable to find object. May be I have designed the model incorrectly, is there any advise...please

Saleh
  

[Li yang Ku]

The object in both the training image and the testing image has to be big enough, otherwise there wont be enough SURF points to recognize. I recommend to start with the model I provided and then move on to something easy like books with text on the cover. Note that this method would not work well on texture less objects. Let me know if you still have problems.

[Saleh]

Thanks Li yang Ku for quick responce.
I had applied your algorithm to detect a pen rather than a drill,
unfortunatly it has failed.
what are the conditions to design the Model?
Best Regards
Saleh

[Li yang Ku]

A pen probably won't work, unless there are words on the pen and you always show the same side of the pen in the test image. Remind that this method uses SIFT like features. SIFT works well on unique patterns with certain size. Some object with a large surface and some special patterns like words will probably work well.

[Saleh]

Mr. Li Yang Ku, could you tell me how can I indicate matching degree. for example: when the algorithm fail to detect the object; I would like to know what was the percentage of matching which based on that percent judged and displayed('object not found')?

Best Regards

Saleh

[Li yang Ku]

Unfortunately this is not a probabilistic method, therefore unlike http://computervisionblog.wordpress.com/2011/12/11/object-recognition/ it doesn't determine an object exist or not base on a probabilistic threshold. However, you should be able to get a degree of confidence based on "length(filter_match)" which is the number of votes remained after filtering. To lower the threshold of finding an object or not, you can modify the parameters I manually set under %check location, %check orientation, %check scale, etc.

[Shikhar]

Hi Mr. Li,

To what extend, will this code work in my requirement.
http://www.mathworks.com/matlabcentral/newsreader/view_thread/319690#875552

Best Regards.