Talk:Essential matrix

I am trying

I am trying to implement an algorithm for calculation of 3D coordinates from corresponding image coordinates pair and I doesn't work fine. I am receiving different results when starting with expresions for x and y. This is what I am doing: 1. Stereo calibrate my setup 2. Inverse camera matrix for both cameras 3. Calculate the normalized coordinates of corresponding points for both cameras - by multipication of inversed camera matrix and homogenous pixel point coordinates 4. Calcule x3 physical/real world coordinate of the point using Rotation and Translation matrices (given by stereo calibration procedure). 5. Calculating x1 and x2 coordinates as it is written in the article

Are you sure there are no mistakes in subsection about 3D poin coordinates? — Preceding unsigned comment added by 194.29.140.25 (talk) 11:48, 7 June 2011 (UTC)

Wrong formula

I think (but I am not positive), that the formula for translation-from-essential is [t] = UZU' instead of VZV'. Could someone smarter check? 2001:718:1E03:5166:F2DE:F1FF:FE7E:60BC (talk) 19:22, 6 November 2013 (UTC)

I entirely agree, http://cseweb.ucsd.edu/classes/sp03/cse252/MaSKS_Ch5.pdf this document on page 4 shows the formula for [t]x = URZU'. Moreover, I have been implementing it with openCV and the results add up with this correction and doesn't with V. — Preceding unsigned comment added by 176.186.116.116 (talk) 18:37, 11 February 2014 (UTC)

No requirement to be normalized points.

"if \mathbf{y} and \mathbf{y}' are homogeneous normalized image coordinates"

There's no requirement that the points be normalized. This is trivial true as if \mathbf{y} is not normalized, then it is a linear multiple of some corresponding normalized point. i.e. a * \mathbf{y}. So the equation y' * E * y = 0 becomes a * a * y' * E * y = 0 which allows both sides to be divided a * a. — Preceding unsigned comment added by 124.168.160.34 (talk) 20:04, 30 July 2014 (UTC)

The term normalized image coordinate does not mean that the corresponding vector of homogeneous coordianates has unit length in R^3. Instead, it is given a precise meaning here, i.e., they refer to a coordinate system in the image plane that has some unit length AND the focal length (distance from image plane to the camera center) is one such length unit. KYN (talk) 18:23, 31 July 2014 (UTC)

Essential from R and T

I have come across sources that show $E=([t]_{x})'R$ instead of $E=R[t]_{x}$ (note transpose in first version). E.g. here http://www.ecse.rpi.edu/~qji/CV/cvfinalproject.pdf (page 2) and here http://cseweb.ucsd.edu/classes/wi05/cse252a/lec14.pdf (page 3, second slide). Also in my experience the former version works in practice while the later does not. — Preceding unsigned comment added by Scuac (talk • contribs) 20:42, 4 September 2014 (UTC)

The formulation of E is highly dependent on how we define the rotation R and translation t. Either they refer to the rotation and translation that transform the coordinate system of the first camera to that of the second one, or vice versa. Depending on which of the two transformations you specify (one is the inverse of the other) you get different expressions for E. --KYN (talk) 14:37, 8 September 2014 (UTC)

That makes perfect sense. I would only point out then that the article makes no clarification on exactly what R is a rotation of. It may be obvious to readers highly familiar with these type of transformations but not to the general reader. May be we can modify the sentence where R is first mentioned like "... where R is a $3\times 3$ rotation matrix from A to B ...", substitute A and B appropriately. --Scuac (talk) 16:06, 8 September 2014 (UTC)