Talk:Lucas–Kanade method

Untitled

There is very little explanation of what is going on before diving in to higher level math. There should be more explanation of what is happening.—WAvegetarian•(talk) 06:21, 27 January 2007 (UTC)

Also it is pretty misleading to do it in 3D - the only case where those equations really apply is if you are tracking something in a CT or MRI scanner. Usually optical flow is 2D. 155.198.65.29 (talk) 09:46, 4 April 2008 (UTC)

The original paper never uses the words "optical flow". This article presents the method in terms of optical flow, which is in fact a separate topic with its own article. Suggest forming the discussion around the concept of disparity and correspondence (as in the original applications to image registration and stereo depth estimation) and moving optical flow to another section. 09:53, 25 June 2008

"As a pixel at location (x, y,z, t) with intensity I(x, y,z, t)" is completely wrong. Pixels do not have 3 Dimensions... should be (x,y,t) in this case. Or change to Voxel. In any case, it is a bit retarded to describe this problem in 3D. Would make it a tad bit easier to understand if you restrict the problem to x,y and t. ~Skela , 31July2008

Implementation issues

As you can read here: http://www.ces.clemson.edu/~stb/klt/ this method has been presented as: An Iterative Image Registration Technique with an Application to Stereo Vision. So why does the article say that it is a non-iterative method? And the reference implementation offered in the same page works with raster images, so why does the articles try to explain it in terms of voxel?. Also, I understand that an important step in the algorithm involves the Newton method. So why the article talks about least squares? Whoever wrote this supposed explanation was on crack. -- Femmina (talk) 16:18, 17 March 2009 (UTC)

LK can be implemented non-iteratively using the least squares method. It is true that many implement it in an iterative scheme, but that is not necessary. On the other hand, some algorithms can only be implemented iteratively. In that sense LK is non-iterative. The implementation given in the article is non-iterative. Do you see any iterativeness in the implementation given in this article?

The article was written for a 3D+t implementation and thus used "voxels". Again, most implementation are 2D+t but that is not always necessary. We can use pixels instead.

This article http://www.mia.uni-saarland.de/Publications/bruhn-ijcv05c.pdf uses least squares for implementing LK (see page 227). Also this http://bmf.hu/conferences/sisy2006/50_Prohaszka.pdf . And this http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.42.5313 one uses non-iterative and least squares method to calculate LK optical flow. This was followed here. Dawoodmajoka (talk) 17:51, 17 March 2009 (UTC)

I didn't check all the links you've given yet, but I've checked out the first, the one titled "Lucas/Kanade Meets Horn/Schunck" and I don't understand why you have to use a questionable research paper as a source for this article. The reference web pages, paper and implementation are available and should be used instead as the primary source and I don't care if you think that KLT can be resolved non iteratively and/or applied to voxels, chickens or whatever. Also, as somebody else already noticed, the words optical flow are used nowhere in the reference material. It's true that is possible to obtain some kind of vector field using a principle derived from the paper (as in the cvCalcOpticalFlowLK function in OpenCV) but that's not in the paper itself. -- Femmina (talk) 23:05, 17 March 2009 (UTC)

"A questionable research paper?" Now you apparently have missed the research on this filed. This paper was awarded the first prize in the conference. AND IJCV is not a "questionable" journal! As far as "optical flow" is concerned, I care even less if those words are used in that paper or not! LK optical flow is an establish method of estimating optical flow. It is irrelevant if the original paper was intended for this application or not. Today, it is used for optical flow estimation too, and everybody in the field knows this. Just google >> optical flow lucas kanade << and you will find numerous papers on this.Dawoodmajoka (talk) 11:32, 18 March 2009 (UTC)

I did google "optical flow lucas kanade" and the first result is this same article as you've written it. The second result I've got is a PDF about a lecture and the fourth is the paper you've already linked here. Doesn't that sound strange to you? And you didn't reply to my question. Why some research papers written by 3rd parties should superceed the reference material published by KLT's authors as the main source and guide for this article? -- Femmina (talk) 12:13, 18 March 2009 (UTC)

I have told you already, that the original paper by LK was on another subject (stereo disparity). Thus it is a futile exercise to demand same implementation here. The basic idea of locally constant flow is decisive. The method as written in this article is same in all papers on LK optical flow upto the solution with the least squares method. It is not an original research. The equations can be solved using various methods, least squares or using newton etc. I am amending the section accordingly and seperating the math details of the method from the least squares implementation. Dawoodmajoka (talk) 19:46, 18 March 2009 (UTC)

You didn't even read the reference paper don't you? -- Femmina (talk) 21:08, 18 March 2009 (UTC)

Did you notice that this article is on the optical flow method called LK (rightly or wrongly) and not about the original article by lucas and kanade which was on a disparity problem.

Detailed treatment

I've removed the Detailed treatment section because it's misleading, contradicts itself, the reference material and contains original research and dubious claims. Various problems with that section have been noticed by users both in this talk page and on Dawoodmajoka's talk page which seems to be the only author. -- Femmina (talk) 18:14, 18 March 2009 (UTC)

I have reverted your changes. This is no way of dealing with diffeence of opinion. You have not presented any alternative, neither have you proposed any improvements. Just to delete something because it does not confrms to your idea of something is not to be made a habit. The only objection to the article so far was about 3d+t implementation vs. 2d+t implementation. At the request and insistence of some readers I have changed to 2d+t. There was no other objection to it. What "dubious claims". If you do not know the field then please refrain from writing unsubstantiated statements! If you have any constructive criticism on the now amendend article, please do share it. Dawoodmajoka (talk) 19:46, 18 March 2009 (UTC)

There is no possible constructive criticism I can offer you for something so badly written. Sorry but the consensus seems to be against your supposed explanation. Issues have been raised by various users and you've failed to take account of any of those while at the same time reverting corrections such as this one: [1] without even considering. -- Femmina (talk) 20:48, 18 March 2009 (UTC)

Offer your "better " version. I am open to any improvements .Vandalism is to delete without offering any improvements. If you do not agree, offer something better. Deletion is no solution. Therefore, I have to revert once again. Almost all criticism so far was on 3d+t implementation, I considered it valid and have switched to 2d+t now. I have also accomodated your criticm on iterative vs. non-iterative and the least-squares implementation. Shows my willingness to compromise. So come up with new criticim or a better discription. Dawoodmajoka (talk) 20:23, 19 March 2009 (UTC)

BTW: I wont be coming available here for some time. Dawoodmajoka (talk) 19:48, 19 March 2009 (UTC)

It's already much better now. Thank you very much for what you've done. However there are still some important aspects missing. The iterative vs non-iterative debate as I understand it revolves around the fact that the concept of optical flow was maybe developed after KLT has been published. What you describe in the article is a method for estimating the optical flow based on some of the concepts published in the paper but there's much more in it. There's a feature detector and an iterative method for estimating the motion of a particular feature between two frames. In my opinion the most interesting part of KLT is the feature tracker and I would like to see it treated in detail rather than the optical flow aspect that, to me, is a bit fuzzy as a concept anyway. -- Femmina (talk) 21:49, 19 March 2009 (UTC)

• "the concept of optical flow was maybe developed after KLT has been published." you are correct. The LK optical flow was devloped after this article was published.
• "What you describe in the article is a method for estimating the optical flow based on some of the concepts published in the paper but there's much more in it." Correct.
• "the most interesting part of KLT is the feature tracker and I would like to see it treated in detail rather than the optical flow aspect": I disagree, that should be in an article on disparity (for which it was originally developed) or in an article on feature tracking etc. This article deals with optical flow LK only. Proposal: Change the title of the article to Lucas-Kanade-Optical-Flow-Method. Dawoodmajoka (talk) 22:34, 19 March 2009 (UTC)

I agree on two different articles because as it is now if I search for "KLT" on wikipedia I land on the Principal components analysis article and a disambiguation message directs me here. If I do the same on google I get Kanade-Lucas-Tomasi Feature Tracker as the first result. So (and that's what actually happened to me) I came to this article hoping to find more informations about the feature tracker and I found the method for estimating the optical flow instead. And I got really confused at first. -- Femmina (talk) 22:48, 19 March 2009 (UTC)

Confusion: Lucas-Kanade vs KLT

Large parts of the confusion in the comments here seem to be due to KLT linking here, which is misleading, as it is not the same. The Kanade-Lucas-Tomasi (KLT) method is a method for tracking image features (see link and link). The foundation of KLT feature tracking and the Lucas-Kanade method for optical flow was laid by the same paper by Lucas and Kanade. Note, however, that this original paper is neither about optical flow nor feature tracking but about registering two images. link

Imho, KLT feature tracking deserves its own article! I don't dare writing it myself at the moment, though. --Sbstn (talk) 17:25, 25 August 2009 (UTC)

Agreed. KLT should not link here. The original Lucas-Kanade paper was about image registration, Tomasi-Kanade added feature selection and tracking and Shi and Tomasi showed the benefit of adding an affine check. This is "KLT tracking" and none of these 3 papers showed derivation of optical flow as formulated on this page. Kegon (talk) 13:53, 13 October 2009 (UTC)

OK, I'm starting a page on KLT tracking Kanade-Lucas-Tomasi Feature Tracker, please come and help get it up to a high standard. Kegon (talk) 14:52, 15 October 2009 (UTC)

External links modified

Hello fellow Wikipedians,

I have just modified one external link on Lucas–Kanade method. Please take a moment to review my edit. If you have any questions, or need the bot to ignore the links, or the page altogether, please visit this simple FaQ for additional information. I made the following changes:

• Added archive https://web.archive.org/web/20140506113705/http://onera-windtunnel.com/dtim-en/gpu-for-image/folkigpu.php to http://onera-windtunnel.com/dtim-en/gpu-for-image/folkigpu.php

When you have finished reviewing my changes, you may follow the instructions on the template below to fix any issues with the URLs.

This message was posted before February 2018. After February 2018, "External links modified" talk page sections are no longer generated or monitored by InternetArchiveBot. No special action is required regarding these talk page notices, other than regular verification using the archive tool instructions below. Editors have permission to delete these "External links modified" talk page sections if they want to de-clutter talk pages, but see the RfC before doing mass systematic removals. This message is updated dynamically through the template {{source check}} (last update: 18 January 2022).

• If you have discovered URLs which were erroneously considered dead by the bot, you can report them with this tool.
• If you found an error with any archives or the URLs themselves, you can fix them with this tool.

Cheers.—InternetArchiveBot (Report bug) 01:35, 8 January 2018 (UTC)

Pronounciation

The IPA reading doesn't seem right to me. I think it should end in e instead of ɪ. Japanese phonology also says that the Japanese e is an mid front unrounded vowel, and not a Near-close near-front unrounded vowel, which seem like an American mispronounciation. If no citation is added, I will remove the pronounciation in a week. --129.247.247.240 (talk) 01:02, 26 March 2018 (UTC)