4

I am trying to implement the algorithm by Jason Hipp et al. There is also a presentation, which is shorter and more comprehensive.

A brief description of their approach:

They use Vector Quantization as a tool to distinguish between foreground and backgroud in any given image. However, instead of using square regions as feature vectors to generate the Codewords, they use circles. This is supposed to decrease the computational complexity. With a circle as predicate vector, the matching problem is reduced to a linear pattern matching task and allows for spatially invariant matching. Hence the method is called Spatially Invariant Vector Quantization.

So basically, a predicate vector is chosen interactively and then the image space is queried exhaustively for the correlation of this predicate vector with the current position.

My questions are:

  • Where in the whole algorithm do they generate the Codebook? And how?

  • I cannot see how to choose the parameters for a Codebook to be generated. If they sample all possible circles in all possible positions in the image first, this is computationally extremely heavy. How do they determine the number of clusters/codewords to be generated?

  • Why would I wobble the sub-rings against each other?

Right now my implementation basically includes one circle with one radius as a predicate vector. It marches through the native image space and correlates the predicate vector with the circle around the current pixel in all possible rotations. This is an extremely slow process and I cannot see the benefits from their algorithm. I have not implemented anything that comes close to a Vector Quantization because I cannot see how this would work.

Any hint or thought is appreciated. The authors of the method didn't respond to my questions, unfortunately.

Tyler Durden
  • 11,156
  • 9
  • 64
  • 126
e0571302
  • 116
  • 6
  • It's a bad article. They don't actually explain the algorithm, they make some hand-waving about how it works and spend most of the time on the examples. It wouldn't pass muster in an image-processing journal. – eh9 Dec 10 '12 at 22:55
  • I agree. Nevertheless, their results are very promising, so I am trying to figure out the method anyway. – e0571302 Dec 11 '12 at 10:14
  • Find the source code they claim in the article. It looks like reading that will be necessary to know what they've done. Perhaps the group has published elsewhere. – eh9 Dec 11 '12 at 14:38
  • Right, its a "concept" article designed to get customers for their services, not a scholarly description of a real algorithm. – Tyler Durden Dec 12 '12 at 15:10
  • Have you had any success with this algorithm? I'm trying to figure whether it's worth spending time on... I agree that their results looked promising. – ojy Feb 03 '15 at 22:40
  • NO @ojy, I haven't had any success, since the major questions were left unanswered - the paper gives no information about it and the authors newer replied to my messages. – e0571302 Mar 30 '15 at 16:36

1 Answers1

2

Your first two questions are not particular to this algorithm, but any vector quantization algorithm. Here is a web page that describes in relatively easy-to-understand terms how to do vector quantization, including generation of codebooks: http://www.data-compression.com/vq.html.

About Wobble: In this algorithm the key observation is that by vectorizing as rings the surface will not be tessellated (fully covered). For example, if you use squares, they tesselate the surface (completely cover it). Overlapping rings will not fully cover the image necessarily. For this reason, pixels which are "between" rings can get missed and cause a failure to match. To compensate for this, the author "wobbles" the rings back and forth so that eventually all the pixels get covered.

Tyler Durden
  • 11,156
  • 9
  • 64
  • 126
  • Thank you Tyler. I understand the principle of Vector Quantization now and don't think I'd have troubles implementing it with square/rectangular regions as feature vectors. – e0571302 Dec 11 '12 at 08:41
  • My concern with rings as feature vectos is actually less the fact that the surface will not be tessellated - but that it will be over"sampled". Every image pixel will be part of a number of rings, which seems highly uneffective. – e0571302 Dec 11 '12 at 09:55
  • Plus, I would need information about the ring radii to sample the surface accordingly. However, I only get this information AFTER the predicate vector has been initialised (otherwise I have to sample the surface with all possible radii). So this is how I don't see how this algorithm could work "on-line" and without prior knowledge of the ring parameters.

    I think I understand your point about the wobbling, thank you for the input!!

     – e0571302 Dec 11 '12 at 09:55