I am developing an application where I am using SIFT + RANSAC and Homography to find an object (OpenCV C++,Java). The problem I am facing is that where there are many outliers RANSAC performs poorly.
For this reasons I would like to try what the author of SIFT said to be pretty good: voting.
I have read that we should vote in a 4 dimension feature space, where the 4 dimensions are:
Location [x, y] (someone says Traslation)
Scale
Orientation
While with opencv is easy to get the match scale and orientation with:
cv::Keypoints.octave
cv::Keypoints.angle
I am having hard time to understand how I can calculate the location.
I have found an interesting slide where with only one match we are able to draw a bounding box:
But I don't get how I could draw that bounding box with just one match. Any help?
You are looking for the largest set of matched features that fit a geometric transformation from image 1 to image 2. In this case, it is the similarity transformation, which has 4 parameters: translation (dx, dy), scale change ds, and rotation d_theta.
Let's say you have matched to features: f1 from image 1 and f2 from image 2. Let (x1,y1) be the location of f1 in image 1, let s1 be its scale, and let theta1 be it's orientation. Similarly you have (x2,y2), s2, and theta2 for f2.
The translation between two features is (dx,dy) = (x2-x1, y2-y1).
The scale change between two features is ds = s2 / s1.
The rotation between two features is d_theta = theta2 - theta1.
So, dx, dy, ds, and d_theta are the dimensions of your Hough space. Each bin corresponds to a similarity transformation.
Once you have performed Hough voting, and found the maximum bin, that bin gives you a transformation from image 1 to image 2. One thing you can do is take the bounding box of image 1 and transform it using that transformation: apply the corresponding translation, rotation and scaling to the corners of the image. Typically, you pack the parameters into a transformation matrix, and use homogeneous coordinates. This will give you the bounding box in image 2 corresponding to the object you've detected.
When using the Hough transform, you create a signature storing the displacement vectors of every feature from the template centroid (either (w/2,h/2) or with the help of central moments).
E.g. for 10 SIFT features found on the template, their relative positions according to template's centroid is a vector<{a,b}>. Now, let's search for this object in a query image: every SIFT feature found in the query image, matched with one of template's 10, casts a vote to its corresponding centroid.
votemap(feature.x - a*, feature.y - b*)+=1 where a,b corresponds to this particular feature vector.
If some of those features cast successfully at the same point (clustering is essential), you have found an object instance.
Signature and voting are reverse procedures. Let's assume V=(-20,-10). So during searching in the novel image, when the two matches are found, we detect their orientation and size and cast a respective vote. E.g. for the right box centroid will be V'=(+20*0.5*cos(-10),+10*0.5*sin(-10)) away from the SIFT feature because it is in half size and rotated by -10 degrees.
To complete Dima's , one needs to add that the 4D Hough space is quantized into a (possibly small) number of 4D boxes, where each box corresponds to the simiéarity given by its center.
Then, for each possible similarity obtained via a tentative matching of features, add 1 into the corresponding box (or cell) in the 4D space. The output similarity is given by the cell with the more votes.
In order to computethe transform from 1 match, just use Dima's formulas in his answer. For several pairs of matches, you may need to use some least squares fit.
Finally, the transform can be applied with the function cv::warpPerspective(), where the third line of the perspective matrix is set to [0,0,1].
Related
Hi i am a beginner in computer vision and i wish to know what exactly is the difference between a homography and affine tranformation, if you want to find the translation between two images which one would you use and why?. From papers and definitions I found online, I am yet to find the difference between them and where one is used instead of the other.
Thanks for your help.
A picture is worth a thousand words:
I have set it down in the terms of a layman.
Homography
A homography, is a matrix that maps a given set of points in one image to the corresponding set of points in another image.
The homography is a 3x3 matrix that maps each point of the first image to the corresponding point of the second image. See below where H is the homography matrix being computed for point x1, y1 and x2, y2
Consider the points of the images present below:
In the case above, there are 4 homography matrices generated.
Where is it used?
You may want to align the above depicted images. You can do so by using the homography.
Here the second image is mapped with respect to the first
Another application is Panoramic Stitching
Visit THIS BLOG for more
Affine transformation
An affine transform generates a matrix to transform the image with respect to the entire image. It does not consider certain points as in the case of homography.
Hence in affine transformation the parallelism of lines is always preserved (as mentioned by EdChum ).
Where is it used?
It is used in areas where you want to alter the entire image:
Rotation (self understood)
Translation (shifting the entire image by a certain length either to top/bottom or left/right)
Scaling (it is basically shrinking or blowing up an image)
See THIS PAGE for more
let's say I am placing a small object on a flat floor inside a room.
First step: Take a picture of the room floor from a known, static position in the world coordinate system.
Second step: Detect the bottom edge of the object in the image and map the pixel coordinate to the object position in the world coordinate system.
Third step: By using a measuring tape measure the real distance to the object.
I could move the small object, repeat this three steps for every pixel coordinate and create a lookup table (key: pixel coordinate; value: distance). This procedure is accurate enough for my use case. I know that it is problematic if there are multiple objects (an object could cover an other object).
My question: Is there an easier way to create this lookup table? Accidentally changing the camera angle by a few degrees destroys the hard work. ;)
Maybe it is possible to execute the three steps for a few specific pixel coordinates or positions in the world coordinate system and perform some "calibration" to calculate the distances with the computed parameters?
If the floor is flat, its equation is that of a plane, let
a.x + b.y + c.z = 1
in the camera coordinates (the origin is the optical center of the camera, XY forms the focal plane and Z the viewing direction).
Then a ray from the camera center to a point on the image at pixel coordinates (u, v) is given by
(u, v, f).t
where f is the focal length.
The ray hits the plane when
(a.u + b.v + c.f) t = 1,
i.e. at the point
(u, v, f) / (a.u + b.v + c.f)
Finally, the distance from the camera to the point is
p = √(u² + v² + f²) / (a.u + b.v + c.f)
This is the function that you need to tabulate. Assuming that f is known, you can determine the unknown coefficients a, b, c by taking three non-aligned points, measuring the image coordinates (u, v) and the distances, and solving a 3x3 system of linear equations.
From the last equation, you can then estimate the distance for any point of the image.
The focal distance can be measured (in pixels) by looking at a target of known size, at a known distance. By proportionality, the ratio of the distance over the size is f over the length in the image.
Most vision libraries (including opencv) have built in functions that will take a couple points from a camera reference frame and the related points from a Cartesian plane and generate your warp matrix (affine transformation) for you. (some are fancy enough to include non-linearity mappings with enough input points, but that brings you back to your time to calibrate issue)
A final note: most vision libraries use some type of grid to calibrate off of ie a checkerboard patter. If you wrote your calibration to work off of such a sheet, then you would only need to measure distances to 1 target object as the transformations would be calculated by the sheet and the target would just provide the world offsets.
I believe what you are after is called a Projective Transformation. The link below should guide you through exactly what you need.
Demonstration of calculating a projective transformation with proper math typesetting on the Math SE.
Although you can solve this by hand and write that into your code... I strongly recommend using a matrix math library or even writing your own matrix math functions prior to resorting to hand calculating the equations as you will have to solve them symbolically to turn it into code and that will be very expansive and prone to miscalculation.
Here are just a few tips that may help you with clarification (applying it to your problem):
-Your A matrix (source) is built from the 4 xy points in your camera image (pixel locations).
-Your B matrix (destination) is built from your measurements in in the real world.
-For fast recalibration, I suggest marking points on the ground to be able to quickly place the cube at the 4 locations (and subsequently get the altered pixel locations in the camera) without having to remeasure.
-You will only have to do steps 1-5 (once) during calibration, after that whenever you want to know the position of something just get the coordinates in your image and run them through step 6 and step 7.
-You will want your calibration points to be as far away from eachother as possible (within reason, as at extreme distances in a vanishing point situation, you start rapidly losing pixel density and therefore source image accuracy). Make sure that no 3 points are colinear (simply put, make your 4 points approximately square at almost the full span of your camera fov in the real world)
ps I apologize for not writing this out here, but they have fancy math editing and it looks way cleaner!
Final steps to applying this method to this situation:
In order to perform this calibration, you will have to set a global home position (likely easiest to do this arbitrarily on the floor and measure your camera position relative to that point). From this position, you will need to measure your object's distance from this position in both x and y coordinates on the floor. Although a more tightly packed calibration set will give you more error, the easiest solution for this may simply be to have a dimension-ed sheet(I am thinking piece of printer paper or a large board or something). The reason that this will be easier is that it will have built in axes (ie the two sides will be orthogonal and you will just use the four corners of the object and used canned distances in your calibration). EX: for a piece of paper your points would be (0,0), (0,8.5), (11,8.5), (11,0)
So using those points and the pixels you get will create your transform matrix, but that still just gives you a global x,y position on axes that may be hard to measure on (they may be skew depending on how you measured/ calibrated). So you will need to calculate your camera offset:
object in real world coords (from steps above): x1, y1
camera coords (Xc, Yc)
dist = sqrt( pow(x1-Xc,2) + pow(y1-Yc,2) )
If it is too cumbersome to try to measure the position of the camera from global origin by hand, you can instead measure the distance to 2 different points and feed those values into the above equation to calculate your camera offset, which you will then store and use anytime you want to get final distance.
As already mentioned in the previous answers you'll need a projective transformation or simply a homography. However, I'll consider it from a more practical view and will try to summarize it short and simple.
So, given the proper homography you can warp your picture of a plane such that it looks like you took it from above (like here). Even simpler you can transform a pixel coordinate of your image to world coordinates of the plane (the same is done during the warping for each pixel).
A homography is basically a 3x3 matrix and you transform a coordinate by multiplying it with the matrix. You may now think, wait 3x3 matrix and 2D coordinates: You'll need to use homogeneous coordinates.
However, most frameworks and libraries will do this handling for you. What you need to do is finding (at least) four points (x/y-coordinates) on your world plane/floor (preferably the corners of a rectangle, aligned with your desired world coordinate system), take a picture of them, measure the pixel coordinates and pass both to the "find-homography-function" of your desired computer vision or math library.
In OpenCV that would be findHomography, here an example (the method perspectiveTransform then performs the actual transformation).
In Matlab you can use something from here. Make sure you are using a projective transformation as transform type. The result is a projective tform, which can be used in combination with this method, in order to transform your points from one coordinate system to another.
In order to transform into the other direction you just have to invert your homography and use the result instead.
I do have two sets of points and I want to find the best transformation between them.
In OpenCV, you have the following function:
Mat H = Calib3d.findHomography(src_points, dest_points);
that returns you a 3x3 Homography matrix, using RANSAC. My problem is now, that I only need translation and rotation (& maybe scale), I don't need affine and perspective.
The thing is, my points are only in 2D.
(1) Is there a function to compute something like a homography but with less degrees of freedom?
(2) If there is none, is it possible to extract a 3x3 matrix that does only translation and rotation from the 3x3 homography matrix?
Thanks in advance for any help!
Isa
OpenCV estimateRigidTransform function is exactly what you need: it returns Translation, Rotation and Scale (use false value for fullAffine flag). And it DOES use RANSAC (see source code to be sure of it).
Homography is for 2D points, the third dimension is just for casting points in 3 dim homogeneous coordinates and performing perspective effects. You can always cast points back:
homogeneous [x, y, w]
cartesian [x/w, y/w]
However since you calculate 6DOF instead of 4DOF (similarity) you result is pretty different from what you expect with 4DOF. More flexible transformation will fit more points in RANSAC at the expense of distortions in transformations you care about. Bottom line - don’t try to decompose H, instead fit similarity or isometry (also called rigid or euclidean). The reason why they are absent in the library - they are expressed in closed form even with correct least squared metric in point coordinates and thus don't require non-linear optimization. In other words, they are very simple.
If you only have rotation and translation, I wrote a quick functions to find them (no RANSAC though). It is probably similar to a rigidTransform but more understandable (hopefully)
https://stackoverflow.com/a/18091472/457687
With scale there is still a closed form solution, but slightly different formulas for translation and scaling. See Learning similarity parameters, p. 25
I am not able to under stand the formula ,
What is W (window) and intensity in the formula mean,
I found this formula in opencv doc
http://docs.opencv.org/trunk/doc/py_tutorials/py_feature2d/py_features_harris/py_features_harris.html
For a grayscale image, intensity levels (0-255) tells you how bright is the pixel..hope that you already know about it.
So, now the explanation of your formula is below:
Aim: We want to find those points which have maximum variation in terms of intensity level in all direction i.e. the points which are very unique in a given image.
I(x,y): This is the intensity value of the current pixel which you are processing at the moment.
I(x+u,y+v): This is the intensity of another pixel which lies at a distance of (u,v) from the current pixel (mentioned above) which is located at (x,y) with intensity I(x,y).
I(x+u,y+v) - I(x,y): This equation gives you the difference between the intensity levels of two pixels.
W(u,v): You don't compare the current pixel with any other pixel located at any random position. You prefer to compare the current pixel with its neighbors so you chose some value for "u" and "v" as you do in case of applying Gaussian mask/mean filter etc. So, basically w(u,v) represents the window in which you would like to compare the intensity of current pixel with its neighbors.
This link explains all your doubts.
For visualizing the algorithm, consider the window function as a BoxFilter, Ix as a Sobel derivative along x-axis and Iy as a Sobel derivative along y-axis.
http://docs.opencv.org/doc/tutorials/imgproc/imgtrans/sobel_derivatives/sobel_derivatives.html will be useful to understand the final equations in the above pdf.
Assuming that I have a grayscale (8-bit) image and assume that I have an integral image created from that same image.
Image resolution is 720x576. According to SURF algorithm, each octave is composed of 4 box filters, which are defined by the number of pixels on their side. The
first octave uses filters with 9x9, 15x15, 21x21 and 27x27 pixels. The
second octave uses filters with 15x15, 27x27, 39x39 and 51x51 pixels.The third octave uses filters with 27x27, 51x51, 75x75 and 99x99 pixels. If the image is sufficiently large and I guess 720x576 is big enough (right??!!), a fourth octave is added, 51x51, 99x99, 147x147 and 195x195. These
octaves partially overlap one another to improve the quality of the interpolated results.
// so, we have:
//
// 9x9 15x15 21x21 27x27
// 15x15 27x27 39x39 51x51
// 27x27 51x51 75x75 99x99
// 51x51 99x99 147x147 195x195
The questions are:What are the values in each of these filters? Should I hardcode these values, or should I calculate them? How exactly (numerically) to apply filters to the integral image?
Also, for calculating the Hessian determinant I found two approximations:
det(HessianApprox) = DxxDyy − (0.9Dxy)^2 anddet(HessianApprox) = DxxDyy − (0.81Dxy)^2Which one is correct?
(Dxx, Dyy, and Dxy are Gaussian second order derivatives).
I had to go back to the original paper to find the precise answers to your questions.
Some background first
SURF leverages a common Image Analysis approach for regions-of-interest detection that is called blob detection.
The typical approach for blob detection is a difference of Gaussians.
There are several reasons for this, the first one being to mimic what happens in the visual cortex of the human brains.
The drawback to difference of Gaussians (DoG) is the computation time that is too expensive to be applied to large image areas.
In order to bypass this issue, SURF takes a simple approach. A DoG is simply the computation of two Gaussian averages (or equivalently, apply a Gaussian blur) followed by taking their difference.
A quick-and-dirty approximation (not so dirty for small regions) is to approximate the Gaussian blur by a box blur.
A box blur is the average value of all the images values in a given rectangle. It can be computed efficiently via integral images.
Using integral images
Inside an integral image, each pixel value is the sum of all the pixels that were above it and on its left in the original image.
The top-left pixel value in the integral image is thus 0, and the bottom-rightmost pixel of the integral image has thus the sum of all the original pixels for value.
Then, you just need to remark that the box blur is equal to the sum of all the pixels inside a given rectangle (not originating in the top-lefmost pixel of the image) and apply the following simple geometric reasoning.
If you have a rectangle with corners ABCD (top left, top right, bottom left, bottom right), then the value of the box filter is given by:
boxFilter(ABCD) = A + D - B - C,
where A, B, C, D is a shortcut for IntegralImagePixelAt(A) (B, C, D respectively).
Integral images in SURF
SURF is not using box blurs of sizes 9x9, etc. directly.
What it uses instead is several orders of Gaussian derivatives, or Haar-like features.
Let's take an example. Suppose you are to compute the 9x9 filters output. This corresponds to a given sigma, hence a fixed scale/octave.
The sigma being fixed, you center your 9x9 window on the pixel of interest. Then, you compute the output of the 2nd order Gaussian derivative in each direction (horizontal, vertical, diagonal). The Fig. 1 in the paper gives you an illustration of the vertical and diagonal filters.
The Hessian determinant
There is a factor to take into account the scale differences. Let's believe the paper that the determinant is equal to:
Det = DxxDyy - (0.9 * Dxy)^2.
Finally, the determinant is given by: Det = DxxDyy - 0.81*Dxy^2.
Look at page 17 of this document
http://www.sci.utah.edu/~fletcher/CS7960/slides/Scott.pdf
If you made a code for normal Gaussian 2D convolution, just use the box filter as a Gaussian kernel and the input image will be the same original image not integral image. The results from this method will be same with the one you asked.