Talk:Optical flow

Robotics Start‑class Mid‑importance

	This article is within the scope of WikiProject Robotics, a collaborative effort to improve the coverage of Robotics on Wikipedia. If you would like to participate, please visit the project page, where you can join the discussion and see a list of open tasks.RoboticsWikipedia:WikiProject RoboticsTemplate:WikiProject RoboticsRobotics
Start	This article has been rated as Start-class on Wikipedia's content assessment scale.
Mid	This article has been rated as Mid-importance on the project's importance scale.

Some Mild Inaccuracies

As a current researcher in the area I am a bit concerned by some confusion some of my students might have visiting this page so I thought to make a list of things that might be beneficial to investigate or change in the event myself or someone else has time.

The inclusion of the reference to Psychologist James Gibson is extremely tertiary at best. While this would be a possible reference for historical motivation of photometric topics, there is no clear tie to optical flow as it is well understood in the field today. It is also a tad confusing to have a psychological reference followed by the stark contrast of a rather heftily mathematical concept.
The first two references provide poor definitions of optical flow. Papers by Horn and Schunck or Beauchemin and Barron provide better sources for inferring a definition of optical flow. As a particular point of inaccuracy, optical flow specifically does not consider anything about the scene or observer itself. It, by definition, only considers a sequence of images and the motion of its intensity field from one image in a sequence to the next. This can be confused by some papers utilizing physical data as a means of finding this mapping between images (thereby introducing new knowns to better resolve the calculation); however, the end result still tells you nothing about the observer or scene itself. This is confirmed by the last few sentences of the article.
The math here should doubtlessly be phrased in terms of two dimensional (i.e. 2D image sequence) optical flow with a corollary regarding N-dimensional optical flow. The 2D case is much easier to put forward to newcomers and the 3D case is not the standard in optical flow and applications are not limited to a ceiling of three dimensions.
The definition of differential optical flow techniques seems to be a misquote of one of the Barron, Fleet, Beauchemin papers. Specifically "Performance of Optical Flow Techniques" which states: "Differential techniques compute velocity from spatiotemporal derivatives of image intensity or filtered versions of the image (using low-pass or band-pass fillters).�" Indeed Taylor series may be used to find constraints in these methods; however, as this paper points out, they are not a general requirement.
The categorization of different flow techniques seems to lack some cohesion, structuring it more closely to the above paper or a similar one comparing optical flow performance would be more structured and provide cited definitions for each category.
The second sentence in "Uses of Optical Flow" is patently incorrect. All papers considering 3D recovery or 3D motion from a 2D sequence have a clear subdivision between Optical Flow (which is often used as a tool in the process) and the resulting 3D geometry or 3D motion field. Only in the case of a 3D image sequence could this be considered close to true, though the geometry still remains a separate problem.
It would be wise to include a small historical section on Horn Schunck optical flow as it is the root paper of many (in my experience the large majority) of optical flow papers. This would additional be a great place to discuss the evolution of optical flow by adding different constraints to complete the problem (Possibly with mention to Lucas Kanade optical flow, another root paper among research).

If anybody finds additional notes or concerns please add them for consideration here. For the best suite of current and archived Optical Flow papers I generally find the IEEE does rather well.

2D-3D

Optical flow can be 2 or 3 dimensional, it can even be n-dimensional. —Preceding unsigned comment added by Dawoodmajoka (talk • contribs) 16:31, 8 February 2009 (UTC)[reply]

The author specifically states that the optical flow is, essentially, a dense motion field. That would make it a noun. However, both this article and the motion estimation article use the term as a verb. Will DIP please make up their collective minds about this and stop confusing the rest of us. Alternatively, please stop using this redundant term all together. Dhatfield (talk) 07:26, 4 April 2008 (UTC)[reply]

On more careful consideration I think that this should be merged into motion estimation. Any expert volunteers? Dhatfield (talk) 07:34, 4 April 2008 (UTC)[reply]

I would be inclined to keep optical flow and motion estimation as distinct topics. Motion estimation is the term used in image coding, while optical flow is used in animal and machine vision. Optical flow is used by humans (and other animals) in ways that have little to do with the image coding algorithms. Optical flow is usually about recovering 3D structure and 3D egomotion, while motion estimation is a pure 2D concept in typical usage. Dicklyon (talk) 06:06, 5 April 2008 (UTC)[reply]

Thanks for the clarification Dick. Overhaul complete. I'm fairly sure it contains errors, but it's improved. Dhatfield (talk) 18:34, 5 April 2008 (UTC)[reply]

I copyedited your stuff to be more concise, fix misspellings and grammar, etc. Dicklyon (talk) 18:41, 5 April 2008 (UTC)[reply]

Merge with optic flow? —Preceding unsigned comment added by 129.215.58.129 (talk) 06:39, 28 April 2008 (UTC)[reply]

YES! Please discuss at this page to avoid duplicate discussions....

Vastly improved from when I last looked here! If you want more pics, talk to me. Dhatfield (talk) 09:52, 18 June 2008 (UTC)[reply]

Arguing that because optical flow is an important component of animal vision that it must also be an important part of computer vision is a logical fallacy. It is equivalent to asserting that, since flapping wings are essential to animal flight, any flying machine must have flapping wings. 19:06, 2 February 2009 (UTC) —Preceding unsigned comment added by 209.211.131.111 (talk)

2D+t instead of 3D+t

The current text describes the case of 3D+t data, but 2D+t is by far the most common type of data used for optical flow techniques and what is discussed in almost all of the literature. I will change to 2D+t shortly unless someone can motivate why 3D+t is better. --KYN (talk) 07:28, 20 February 2009 (UTC)[reply]

As the author of this part of the article, I have now switched the flow equations to the 2D format. Dawoodmajoka (talk) 20:13, 20 February 2009 (UTC)[reply]

Slightly strange vector maths

Normally the dot product is between two vectors $\mathbf {u} \cdot \mathbf {v}$ , or if you use the transpose notation, a matrix product between a row and a column vector written without a dot: $\mathbf {u} ^{T}\mathbf {v}$ . This article's first section uses *both* of these together, which looks a little odd. —Preceding unsigned comment added by 88.110.199.11 (talk) 21:08, 4 October 2009 (UTC)[reply]

Merge from Optical flow sensor

I think the sensors should be discussed in the same article as the methods; no need to separate the material into two short articles just because one involves hardware. Dicklyon (talk) 17:01, 10 April 2011 (UTC)[reply]

I thought about that as well- about adding a section to "optical flow" regarding computational hardware. However the page for optical flow itself needs updating as well, since the term can refer to both the psychological phenomena as well as the algorithms used to measure it. This subject brings together biology, neuroscience, image processing, robotics, and so on... User:Trevmicro —Preceding undated comment added 18:01, 10 April 2011 (UTC). PS. No objections from me.[reply]

Then do that if you're up for it. Change the new article to a redirect to here; ask if you need help on how. Dicklyon (talk) 19:01, 10 April 2011 (UTC)[reply]