Related
I have two data files, each of them contain a big number of 3-dimensional points (file A stores approximately 50,000 points, file B stores approximately 500,000 points). My goal is to find for every point (a) in file A the point (b) in file B which has the smallest distance to (a). I store the points in two lists like this:
List A nodes:
(ID X Y Z)
[ ['478277', -107.0, 190.5674, 128.1634],
['478279', -107.0, 190.5674, 134.0172],
['478282', -107.0, 190.5674, 131.0903],
['478283', -107.0, 191.9798, 124.6807],
... ]
List B data:
(X Y Z Data)
[ [-28.102, 173.657, 229.744, 14.318],
[-28.265, 175.549, 227.824, 13.648],
[-27.695, 175.925, 227.133, 13.142],
...]
My first approach was to simply iterate through the first and second list with a nested loop and compute the distance between every points like this:
outfile = open(job[0] + '/' + output, 'wb');
dist_min = float(job[5]);
dist_max = float(job[6]);
dists = [];
for node in nodes:
shortest_distance = 1000.0;
shortest_data = 0.0;
for entry in data:
dist = math.sqrt((node[1] - entry[0])**2 + (node[2] - entry[1])**2 + (node[3] - entry[2])**2);
if (dist_min <= dist <= dist_max) and (dist < shortest_distance):
shortest_distance = dist;
shortest_data = entry[3];
outfile.write(node[0] + ', ' + str('%10.5f' % shortest_data + '\n'));
outfile.close();
I recognized that the amount of loops Python has to run is way too big (~25,000,000,000), so I had to fasten my code. I tried to first calculate all distances with list comprehensions but the code still is too slow:
p_x = [row[1] for row in nodes];
p_y = [row[2] for row in nodes];
p_z = [row[3] for row in nodes];
q_x = [row[0] for row in data];
q_y = [row[1] for row in data];
q_z = [row[2] for row in data];
dx = [[(px - qx) for px in p_x] for qx in q_x];
dy = [[(py - qy) for py in p_y] for qy in q_y];
dz = [[(pz - qz) for pz in p_z] for qz in q_z];
dx = [[dxxx * dxxx for dxxx in dxx] for dxx in dx];
dy = [[dyyy * dyyy for dyyy in dyy] for dyy in dy];
dz = [[dzzz * dzzz for dzzz in dzz] for dzz in dz];
D = [[(dx[i][j] + dy[i][j] + dz[i][j]) for j in range(len(dx[0]))] for i in range(len(dx))];
D = [[(DDD**(0.5)) for DDD in DD] for DD in D];
To be honest, at this point, I do not know which of the two approaches is better, anyway, none of the two possibilities seem feasible. I'm not even sure if it is possible to write a code which calculates all distances in an acceptable time. Is there even another way to solve my problem without calculating all distances?
Edit: I forgot to mention that I am running on Python 2.5.1 and am not allowed to install or add any new libraries...
Just in case someone is interrested in the solution:
I found a way to speed up the whole process by not calculating all distances:
I created a 3D-list, representing a grid in the given 3D space, divided in X, Y and Z in a given step size (e.g. (Max. - Min.) / 1,000). Then I iterated over every 3D point to put it into my grid. After that I iterated over the points of set A again, looking if there are points from B in the same cube, if not I would increase the search radius, so the process is looking in the adjacent 26 cubes for points. The radius is increasing until there is at least one point found. The resulting list is comparatively small and can be ordered in short time and the nearest point is found.
The processing time went down to a couple minutes and it is working fine.
p_x = [row[1] for row in nodes];
p_y = [row[2] for row in nodes];
p_z = [row[3] for row in nodes];
q_x = [row[0] for row in data];
q_y = [row[1] for row in data];
q_z = [row[2] for row in data];
min_x = min(p_x + q_x);
min_y = min(p_y + q_y);
min_z = min(p_z + q_z);
max_x = max(p_x + q_x);
max_y = max(p_y + q_y);
max_z = max(p_z + q_z);
max_n = max(max_x, max_y, max_z);
min_n = min(min_x, min_y, max_z);
gridcount = 1000;
step = (max_n - min_n) / gridcount;
ruler_x = [min_x + (i * step) for i in range(gridcount + 1)];
ruler_y = [min_y + (i * step) for i in range(gridcount + 1)];
ruler_z = [min_z + (i * step) for i in range(gridcount + 1)];
grid = [[[0 for i in range(gridcount)] for j in range(gridcount)] for k in range(gridcount)];
for node in nodes:
loc_x = self.abatemp_get_cell(node[1], ruler_x);
loc_y = self.abatemp_get_cell(node[2], ruler_y);
loc_z = self.abatemp_get_cell(node[3], ruler_z);
if grid[loc_x][loc_y][loc_z] is 0:
grid[loc_x][loc_y][loc_z] = [[node[1], node[2], node[3], node[0]]];
else:
grid[loc_x][loc_y][loc_z].append([node[1], node[2], node[3], node[0]]);
for entry in data:
loc_x = self.abatemp_get_cell(entry[0], ruler_x);
loc_y = self.abatemp_get_cell(entry[1], ruler_y);
loc_z = self.abatemp_get_cell(entry[2], ruler_z);
if grid[loc_x][loc_y][loc_z] is 0:
grid[loc_x][loc_y][loc_z] = [[entry[0], entry[1], entry[2], entry[3]]];
else:
grid[loc_x][loc_y][loc_z].append([entry[0], entry[1], entry[2], entry[3]]);
out = [];
outfile = open(job[0] + '/' + output, 'wb');
for node in nodes:
neighbours = [];
radius = -1;
loc_nx = self.abatemp_get_cell(node[1], ruler_x);
loc_ny = self.abatemp_get_cell(node[2], ruler_y);
loc_nz = self.abatemp_get_cell(node[3], ruler_z);
reloop = True;
while reloop:
if neighbours:
reloop = False;
radius += 1;
start_x = 0 if ((loc_nx - radius) < 0) else (loc_nx - radius);
start_y = 0 if ((loc_ny - radius) < 0) else (loc_ny - radius);
start_z = 0 if ((loc_nz - radius) < 0) else (loc_nz - radius);
end_x = (len(ruler_x) - 1) if ((loc_nx + radius + 1) > (len(ruler_x) - 1)) else (loc_nx + radius + 1);
end_y = (len(ruler_y) - 1) if ((loc_ny + radius + 1) > (len(ruler_y) - 1)) else (loc_ny + radius + 1);
end_z = (len(ruler_z) - 1) if ((loc_nz + radius + 1) > (len(ruler_z) - 1)) else (loc_nz + radius + 1);
for i in range(start_x, end_x):
for j in range(start_y, end_y):
for k in range(start_z, end_z):
if not grid[i][j][k] is 0:
for grid_entry in grid[i][j][k]:
if not isinstance(grid_entry[3], basestring):
neighbours.append(grid_entry);
dists = [];
for n in neighbours:
d = math.sqrt((node[1] - n[0])**2 + (node[2] - n[1])**2 + (node[3] - n[2])**2);
dists.append([d, n[3]]);
dists = sorted(dists);
outfile.write(node[0] + ', ' + str(dists[0][-1]) + '\n');
outfile.close();
Function to get the position of a point:
def abatemp_get_cell(self, n, ruler):
for i in range(len(ruler)):
if i >= len(ruler):
return False;
if ruler[i] <= n <= ruler[i + 1]:
return i;
The gridcount variable gives one the chance to fasten the process, with a small gridcount the process of sorting the points into the grid is very fast, but the lists of neighbours in the search loop gets bigger and more time is needed for this part of the process. With a big gridcount more time is needed at the beginning, however the loop runs faster.
The only issue I have now is the fact, that there are cases when the process found neighbours but there are other points, which are not yet found, but are closer to the point (see picture). So far I solved this issue by incrementing the search radius another time when there are already neigbours. And still then I have points which are closer but not in the neighbours list, although it's a very small amount (92 out of ~100,000). I could solve this problem by increment the radius two times after finding neighbours, but this solution seems not very smart. Maybe you guys have an idea...
This is the first working draft of the process, I think it will be possible to improve it even more, just to give you an idea of how it is working...
It took me a bit of thought but at the end I think I found a solution for you.
Your problem is not in the code you wrote but in the algorithm it implements.
There is an algorithm called Dijkstra's algorithm and here is the gist of it: https://en.wikipedia.org/wiki/Dijkstra%27s_algorithm .
Now what you need to do is to use this algorithm in a clever way:
create a node S (stand for source).
Now link edges from S to all the nodes in B group.
After you done that you should link edges from each point b in B to each point a in A.
You should set the cost of the links from the source to 0 and the other to the distance between 2 points (only in 3D).
Now if we will use Dijkstra's algorithm the output we will get would be the cost to travel from S to each point in the graph (we are only interested in the distance to points in group A).
So since the cost is 0 to each point b in B and S is only connected to points in B so the road to any point a in A must include a node in B (actually exactly one since the shortest distance between to points is a single line).
I am not sure if this will fasten your code but as far as I know, a way to solve this problem without calculating all distances does not exist and this algorithm is the best time complexity one could hope for.
take a look at this generic 3D data structure:
https://github.com/m4nh/skimap_ros
it has a very fast RadiusSearch feature just ready to be used. This solution (similar to Octree but faster) avoids to you to create the Regular Grid first (you don't have to fix MAX/MIN size along each axis) and you save a lot of memory
I would like to uniformly distribute a predetermined set of points within a circle. By uniform distribution, I mean they should all be equally distanced from each other (hence a random approach won't work). I tried a hexagonal approach, but I had problems consistently reaching the outermost radius.
My current approach is a nested for loop where each outer iteration reduces the radius & number of points, and each inner loop evenly drops points on the new radius. Essentially, it's a bunch of nested circles. Unfortunately, it's far from even. Any tips on how to do this correctly?
The goals of having a uniform distribution within the area and a uniform distribution on the boundary conflict; any solution will be a compromise between the two. I augmented the sunflower seed arrangement with an additional parameter alpha that indicates how much one cares about the evenness of boundary.
alpha=0 gives the typical sunflower arrangement, with jagged boundary:
With alpha=2 the boundary is smoother:
(Increasing alpha further is problematic: Too many points end up on the boundary).
The algorithm places n points, of which the kth point is put at distance sqrt(k-1/2) from the boundary (index begins with k=1), and with polar angle 2*pi*k/phi^2 where phi is the golden ratio. Exception: the last alpha*sqrt(n) points are placed on the outer boundary of the circle, and the polar radius of other points is scaled to account for that. This computation of the polar radius is done in the function radius.
It is coded in MATLAB.
function sunflower(n, alpha) % example: n=500, alpha=2
clf
hold on
b = round(alpha*sqrt(n)); % number of boundary points
phi = (sqrt(5)+1)/2; % golden ratio
for k=1:n
r = radius(k,n,b);
theta = 2*pi*k/phi^2;
plot(r*cos(theta), r*sin(theta), 'r*');
end
end
function r = radius(k,n,b)
if k>n-b
r = 1; % put on the boundary
else
r = sqrt(k-1/2)/sqrt(n-(b+1)/2); % apply square root
end
end
Might as well tag on my Python translation.
from math import sqrt, sin, cos, pi
phi = (1 + sqrt(5)) / 2 # golden ratio
def sunflower(n, alpha=0, geodesic=False):
points = []
angle_stride = 360 * phi if geodesic else 2 * pi / phi ** 2
b = round(alpha * sqrt(n)) # number of boundary points
for k in range(1, n + 1):
r = radius(k, n, b)
theta = k * angle_stride
points.append((r * cos(theta), r * sin(theta)))
return points
def radius(k, n, b):
if k > n - b:
return 1.0
else:
return sqrt(k - 0.5) / sqrt(n - (b + 1) / 2)
# example
if __name__ == '__main__':
import matplotlib.pyplot as plt
fig, ax = plt.subplots()
points = sunflower(500, alpha=2, geodesic=False)
xs = [point[0] for point in points]
ys = [point[1] for point in points]
ax.scatter(xs, ys)
ax.set_aspect('equal') # display as square plot with equal axes
plt.show()
Stumbled across this question and the answer above (so all cred to user3717023 & Matt).
Just adding my translation into R here, in case someone else needed that :)
library(tibble)
library(dplyr)
library(ggplot2)
sunflower <- function(n, alpha = 2, geometry = c('planar','geodesic')) {
b <- round(alpha*sqrt(n)) # number of boundary points
phi <- (sqrt(5)+1)/2 # golden ratio
r <- radius(1:n,n,b)
theta <- 1:n * ifelse(geometry[1] == 'geodesic', 360*phi, 2*pi/phi^2)
tibble(
x = r*cos(theta),
y = r*sin(theta)
)
}
radius <- function(k,n,b) {
ifelse(
k > n-b,
1,
sqrt(k-1/2)/sqrt(n-(b+1)/2)
)
}
# example:
sunflower(500, 2, 'planar') %>%
ggplot(aes(x,y)) +
geom_point()
Building on top of #OlivelsAWord , here is a Python implementation using numpy:
import numpy as np
import matplotlib.pyplot as plt
def sunflower(n: int, alpha: float) -> np.ndarray:
# Number of points respectively on the boundary and inside the cirlce.
n_exterior = np.round(alpha * np.sqrt(n)).astype(int)
n_interior = n - n_exterior
# Ensure there are still some points in the inside...
if n_interior < 1:
raise RuntimeError(f"Parameter 'alpha' is too large ({alpha}), all "
f"points would end-up on the boundary.")
# Generate the angles. The factor k_theta corresponds to 2*pi/phi^2.
k_theta = np.pi * (3 - np.sqrt(5))
angles = np.linspace(k_theta, k_theta * n, n)
# Generate the radii.
r_interior = np.sqrt(np.linspace(0, 1, n_interior))
r_exterior = np.ones((n_exterior,))
r = np.concatenate((r_interior, r_exterior))
# Return Cartesian coordinates from polar ones.
return r * np.stack((np.cos(angles), np.sin(angles)))
# NOTE: say the returned array is called s. The layout is such that s[0,:]
# contains X values and s[1,:] contains Y values. Change the above to
# return r.reshape(n, 1) * np.stack((np.cos(angles), np.sin(angles)), axis=1)
# if you want s[:,0] and s[:,1] to contain X and Y values instead.
if __name__ == '__main__':
fig, ax = plt.subplots()
# Let's plot three sunflowers with different values of alpha!
for alpha in (0, 1, 2):
s = sunflower(500, alpha)
# NOTE: the 'alpha=0.5' parameter is to control transparency, it does
# not have anything to do with the alpha used in 'sunflower' ;)
ax.scatter(s[0], s[1], alpha=0.5, label=f"alpha={alpha}")
# Display as square plot with equal axes and add a legend. Then show the result :)
ax.set_aspect('equal')
ax.legend()
plt.show()
Adding my Java implementation of previous answers with an example (Processing).
int n = 2000; // count of nodes
Float alpha = 2.; // constant that can be adjusted to vary the geometry of points at the boundary
ArrayList<PVector> vertices = new ArrayList<PVector>();
Float scaleFactor = 200.; // scale points beyond their 0.0-1.0 range for visualisation;
void setup() {
size(500, 500);
// Test
vertices = sunflower(n, alpha);
displayTest(vertices, scaleFactor);
}
ArrayList<PVector> sunflower(int n, Float alpha) {
Double phi = (1 + Math.sqrt(5)) / 2; // golden ratio
Double angle = 2 * PI / Math.pow(phi, 2); // value used to calculate theta for each point
ArrayList<PVector> points = new ArrayList<PVector>();
Long b = Math.round(alpha*Math.sqrt(n)); // number of boundary points
Float theta, r, x, y;
for (int i = 1; i < n + 1; i++) {
r = radius(i, n, b.floatValue());
theta = i * angle.floatValue();
x = r * cos(theta);
y = r * sin(theta);
PVector p = new PVector(x, y);
points.add(p);
}
return points;
}
Float radius(int k, int n, Float b) {
if (k > n - b) {
return 1.0;
} else {
Double r = Math.sqrt(k - 0.5) / Math.sqrt(n - (b+1) / 2);
return r.floatValue();
}
}
void displayTest(ArrayList<PVector> points, Float size) {
for (int i = 0; i < points.size(); i++) {
Float x = size * points.get(i).x;
Float y = size * points.get(i).y;
pushMatrix();
translate(width / 2, height / 2);
ellipse(x, y, 5, 5);
popMatrix();
}
}
Here's my Unity implementation.
Vector2[] Sunflower(int n, float alpha = 0, bool geodesic = false){
float phi = (1 + Mathf.Sqrt(5)) / 2;//golden ratio
float angle_stride = 360 * phi;
float radius(float k, float n, float b)
{
return k > n - b ? 1 : Mathf.Sqrt(k - 0.5f) / Mathf.Sqrt(n - (b + 1) / 2);
}
int b = (int)(alpha * Mathf.Sqrt(n)); //# number of boundary points
List<Vector2>points = new List<Vector2>();
for (int k = 0; k < n; k++)
{
float r = radius(k, n, b);
float theta = geodesic ? k * 360 * phi : k * angle_stride;
float x = !float.IsNaN(r * Mathf.Cos(theta)) ? r * Mathf.Cos(theta) : 0;
float y = !float.IsNaN(r * Mathf.Sin(theta)) ? r * Mathf.Sin(theta) : 0;
points.Add(new Vector2(x, y));
}
return points.ToArray();
}
I have a dynamic range of X by Y sheet sizes which need to be tested by a long running algorithm.
For example, I am going to try any sheet size between 40-50 inches wide by 40-80 inches long. It would take too long to try every integer combination so I want to limit the iterations to 30.
Since there are only 10 units in the X range and 40 units in the Y range, I need to test about 3 X units and 10 Y units and skip the rest.
How can this be coded to figure out the closest ratio and end up with only 30 iterations? It needs to be dynamic because these ranges keep changing and sometimes the Y range is less than the X range.
answer: (inspired by Fraser)
Dim ratioX As Integer = txtSizeFormSingleXmax - txtSizeFormSingleXmin
Dim ratioY As Integer = txtSizeFormSingleYmax - txtSizeFormSingleYmin
Dim FinalRatioNumerator As Integer
Dim FinalRatioDenominator As Integer
Dim XGreaterThanY As Boolean = False
If ratioX > ratioY Then
Dim tempRatio As Integer
tempRatio = ratioY
ratioY = ratioX
ratioX = tempRatio
XGreaterThanY = True
End If
For countRatio As Integer = 1 To 30
If ratioX / ratioY <= countRatio / CInt(30 / countRatio) Then
FinalRatioNumerator = countRatio
FinalRatioDenominator = CInt(30 / countRatio)
Exit For
End If
Next
Since # of iterations must be whole numbers our options are 1,30 2,15 3,10 and 5,6 you may want to consider 4,7 or 4,8 though 30 iterations is not met.
if i pseudo this up without optimization so its easier to follow my logic
get ranges x and y
if x>y y/x = ratio else swap x and y
now all you really need to do (given you are set on 30 ish iterations)
calc difference in ratio
for each pair possible
(((x/y)-(1/30 ))^2)^(1/2) =result
now you just take the closest match (smallest result) and use those
as your test numbers
for each xtestval in mylist of xtestvals
(Xrange / xTestNum) + xrange min = xtestval
mylist.add(xtestval)
xrange min = xtestval
next
then you have your test values for each axis so you double loop to get the combinations
for each x
for each y
pair x and y
next
next
not super clear but hopefully clear enough... back to work for me!
I'm going to rephrase your question, is this decription correct? Given a taret ratio:
W / H
You want to optimize X / Y as close to W / H as possible, while xmin < x < xmax and ymin < y < ymax.
Proposed algorithm:
var x = xmin + xmax / 2;
var y = ymin + ymax / 2;
function int[] adjustRange(x, y, wHRatio)
{
var curRatio = x / y;
if(curRatio == wHRatio)
{
return [x, y];
}
else if(curRatio < wHRatio)
{
// you want to increase x or decrease y. If neither are possible, x/y
// is as close as you will get. You can fudge your intervals to be
// based on the xmax-xmin and ymax-ymin, or some number from a
// config file
}
else if(curRatio > wHRatio)
{
// you want to decrease x or increase y. If neither are possible, x/y
// is as close as you will get
}
}
BOUNTY STATUS UPDATE:
I discovered how to map a linear lens, from destination coordinates to source coordinates.
How do you calculate the radial distance from the centre to go from fisheye to rectilinear?
1). I actually struggle to reverse it, and to map source coordinates to destination coordinates. What is the inverse, in code in the style of the converting functions I posted?
2). I also see that my undistortion is imperfect on some lenses - presumably those that are not strictly linear. What is the equivalent to-and-from source-and-destination coordinates for those lenses? Again, more code than just mathematical formulae please...
Question as originally stated:
I have some points that describe positions in a picture taken with a fisheye lens.
I want to convert these points to rectilinear coordinates. I want to undistort the image.
I've found this description of how to generate a fisheye effect, but not how to reverse it.
There's also a blog post that describes how to use tools to do it; these pictures are from that:
(1) : SOURCE Original photo link
Input : Original image with fish-eye distortion to fix.
(2) : DESTINATION Original photo link
Output : Corrected image (technically also with perspective correction, but that's a separate step).
How do you calculate the radial distance from the centre to go from fisheye to rectilinear?
My function stub looks like this:
Point correct_fisheye(const Point& p,const Size& img) {
// to polar
const Point centre = {img.width/2,img.height/2};
const Point rel = {p.x-centre.x,p.y-centre.y};
const double theta = atan2(rel.y,rel.x);
double R = sqrt((rel.x*rel.x)+(rel.y*rel.y));
// fisheye undistortion in here please
//... change R ...
// back to rectangular
const Point ret = Point(centre.x+R*cos(theta),centre.y+R*sin(theta));
fprintf(stderr,"(%d,%d) in (%d,%d) = %f,%f = (%d,%d)\n",p.x,p.y,img.width,img.height,theta,R,ret.x,ret.y);
return ret;
}
Alternatively, I could somehow convert the image from fisheye to rectilinear before finding the points, but I'm completely befuddled by the OpenCV documentation. Is there a straightforward way to do it in OpenCV, and does it perform well enough to do it to a live video feed?
The description you mention states that the projection by a pin-hole camera (one that does not introduce lens distortion) is modeled by
R_u = f*tan(theta)
and the projection by common fisheye lens cameras (that is, distorted) is modeled by
R_d = 2*f*sin(theta/2)
You already know R_d and theta and if you knew the camera's focal length (represented by f) then correcting the image would amount to computing R_u in terms of R_d and theta. In other words,
R_u = f*tan(2*asin(R_d/(2*f)))
is the formula you're looking for. Estimating the focal length f can be solved by calibrating the camera or other means such as letting the user provide feedback on how well the image is corrected or using knowledge from the original scene.
In order to solve the same problem using OpenCV, you would have to obtain the camera's intrinsic parameters and lens distortion coefficients. See, for example, Chapter 11 of Learning OpenCV (don't forget to check the correction). Then you can use a program such as this one (written with the Python bindings for OpenCV) in order to reverse lens distortion:
#!/usr/bin/python
# ./undistort 0_0000.jpg 1367.451167 1367.451167 0 0 -0.246065 0.193617 -0.002004 -0.002056
import sys
import cv
def main(argv):
if len(argv) < 10:
print 'Usage: %s input-file fx fy cx cy k1 k2 p1 p2 output-file' % argv[0]
sys.exit(-1)
src = argv[1]
fx, fy, cx, cy, k1, k2, p1, p2, output = argv[2:]
intrinsics = cv.CreateMat(3, 3, cv.CV_64FC1)
cv.Zero(intrinsics)
intrinsics[0, 0] = float(fx)
intrinsics[1, 1] = float(fy)
intrinsics[2, 2] = 1.0
intrinsics[0, 2] = float(cx)
intrinsics[1, 2] = float(cy)
dist_coeffs = cv.CreateMat(1, 4, cv.CV_64FC1)
cv.Zero(dist_coeffs)
dist_coeffs[0, 0] = float(k1)
dist_coeffs[0, 1] = float(k2)
dist_coeffs[0, 2] = float(p1)
dist_coeffs[0, 3] = float(p2)
src = cv.LoadImage(src)
dst = cv.CreateImage(cv.GetSize(src), src.depth, src.nChannels)
mapx = cv.CreateImage(cv.GetSize(src), cv.IPL_DEPTH_32F, 1)
mapy = cv.CreateImage(cv.GetSize(src), cv.IPL_DEPTH_32F, 1)
cv.InitUndistortMap(intrinsics, dist_coeffs, mapx, mapy)
cv.Remap(src, dst, mapx, mapy, cv.CV_INTER_LINEAR + cv.CV_WARP_FILL_OUTLIERS, cv.ScalarAll(0))
# cv.Undistort2(src, dst, intrinsics, dist_coeffs)
cv.SaveImage(output, dst)
if __name__ == '__main__':
main(sys.argv)
Also note that OpenCV uses a very different lens distortion model to the one in the web page you linked to.
(Original poster, providing an alternative)
The following function maps destination (rectilinear) coordinates to source (fisheye-distorted) coordinates. (I'd appreciate help in reversing it)
I got to this point through trial-and-error: I don't fundamentally grasp why this code is working, explanations and improved accuracy appreciated!
def dist(x,y):
return sqrt(x*x+y*y)
def correct_fisheye(src_size,dest_size,dx,dy,factor):
""" returns a tuple of source coordinates (sx,sy)
(note: values can be out of range)"""
# convert dx,dy to relative coordinates
rx, ry = dx-(dest_size[0]/2), dy-(dest_size[1]/2)
# calc theta
r = dist(rx,ry)/(dist(src_size[0],src_size[1])/factor)
if 0==r:
theta = 1.0
else:
theta = atan(r)/r
# back to absolute coordinates
sx, sy = (src_size[0]/2)+theta*rx, (src_size[1]/2)+theta*ry
# done
return (int(round(sx)),int(round(sy)))
When used with a factor of 3.0, it successfully undistorts the images used as examples (I made no attempt at quality interpolation):
Dead link
(And this is from the blog post, for comparison:)
If you think your formulas are exact, you can comput an exact formula with trig, like so:
Rin = 2 f sin(w/2) -> sin(w/2)= Rin/2f
Rout= f tan(w) -> tan(w)= Rout/f
(Rin/2f)^2 = [sin(w/2)]^2 = (1 - cos(w))/2 -> cos(w) = 1 - 2(Rin/2f)^2
(Rout/f)^2 = [tan(w)]^2 = 1/[cos(w)]^2 - 1
-> (Rout/f)^2 = 1/(1-2[Rin/2f]^2)^2 - 1
However, as #jmbr says, the actual camera distortion will depend on the lens and the zoom. Rather than rely on a fixed formula, you might want to try a polynomial expansion:
Rout = Rin*(1 + A*Rin^2 + B*Rin^4 + ...)
By tweaking first A, then higher-order coefficients, you can compute any reasonable local function (the form of the expansion takes advantage of the symmetry of the problem). In particular, it should be possible to compute initial coefficients to approximate the theoretical function above.
Also, for good results, you will need to use an interpolation filter to generate your corrected image. As long as the distortion is not too great, you can use the kind of filter you would use to rescale the image linearly without much problem.
Edit: as per your request, the equivalent scaling factor for the above formula:
(Rout/f)^2 = 1/(1-2[Rin/2f]^2)^2 - 1
-> Rout/f = [Rin/f] * sqrt(1-[Rin/f]^2/4)/(1-[Rin/f]^2/2)
If you plot the above formula alongside tan(Rin/f), you can see that they are very similar in shape. Basically, distortion from the tangent becomes severe before sin(w) becomes much different from w.
The inverse formula should be something like:
Rin/f = [Rout/f] / sqrt( sqrt(([Rout/f]^2+1) * (sqrt([Rout/f]^2+1) + 1) / 2 )
I blindly implemented the formulas from here, so I cannot guarantee it would do what you need.
Use auto_zoom to get the value for the zoom parameter.
def dist(x,y):
return sqrt(x*x+y*y)
def fisheye_to_rectilinear(src_size,dest_size,sx,sy,crop_factor,zoom):
""" returns a tuple of dest coordinates (dx,dy)
(note: values can be out of range)
crop_factor is ratio of sphere diameter to diagonal of the source image"""
# convert sx,sy to relative coordinates
rx, ry = sx-(src_size[0]/2), sy-(src_size[1]/2)
r = dist(rx,ry)
# focal distance = radius of the sphere
pi = 3.1415926535
f = dist(src_size[0],src_size[1])*factor/pi
# calc theta 1) linear mapping (older Nikon)
theta = r / f
# calc theta 2) nonlinear mapping
# theta = asin ( r / ( 2 * f ) ) * 2
# calc new radius
nr = tan(theta) * zoom
# back to absolute coordinates
dx, dy = (dest_size[0]/2)+rx/r*nr, (dest_size[1]/2)+ry/r*nr
# done
return (int(round(dx)),int(round(dy)))
def fisheye_auto_zoom(src_size,dest_size,crop_factor):
""" calculate zoom such that left edge of source image matches left edge of dest image """
# Try to see what happens with zoom=1
dx, dy = fisheye_to_rectilinear(src_size, dest_size, 0, src_size[1]/2, crop_factor, 1)
# Calculate zoom so the result is what we wanted
obtained_r = dest_size[0]/2 - dx
required_r = dest_size[0]/2
zoom = required_r / obtained_r
return zoom
I took what JMBR did and basically reversed it. He took the radius of the distorted image (Rd, that is, the distance in pixels from the center of the image) and found a formula for Ru, the radius of the undistorted image.
You want to go the other way. For each pixel in the undistorted (processed image), you want to know what the corresponding pixel is in the distorted image.
In other words, given (xu, yu) --> (xd, yd). You then replace each pixel in the undistorted image with its corresponding pixel from the distorted image.
Starting where JMBR did, I do the reverse, finding Rd as a function of Ru. I get:
Rd = f * sqrt(2) * sqrt( 1 - 1/sqrt(r^2 +1))
where f is the focal length in pixels (I'll explain later), and r = Ru/f.
The focal length for my camera was 2.5 mm. The size of each pixel on my CCD was 6 um square. f was therefore 2500/6 = 417 pixels. This can be found by trial and error.
Finding Rd allows you to find the corresponding pixel in the distorted image using polar coordinates.
The angle of each pixel from the center point is the same:
theta = arctan( (yu-yc)/(xu-xc) ) where xc, yc are the center points.
Then,
xd = Rd * cos(theta) + xc
yd = Rd * sin(theta) + yc
Make sure you know which quadrant you are in.
Here is the C# code I used
public class Analyzer
{
private ArrayList mFisheyeCorrect;
private int mFELimit = 1500;
private double mScaleFESize = 0.9;
public Analyzer()
{
//A lookup table so we don't have to calculate Rdistorted over and over
//The values will be multiplied by focal length in pixels to
//get the Rdistorted
mFisheyeCorrect = new ArrayList(mFELimit);
//i corresponds to Rundist/focalLengthInPixels * 1000 (to get integers)
for (int i = 0; i < mFELimit; i++)
{
double result = Math.Sqrt(1 - 1 / Math.Sqrt(1.0 + (double)i * i / 1000000.0)) * 1.4142136;
mFisheyeCorrect.Add(result);
}
}
public Bitmap RemoveFisheye(ref Bitmap aImage, double aFocalLinPixels)
{
Bitmap correctedImage = new Bitmap(aImage.Width, aImage.Height);
//The center points of the image
double xc = aImage.Width / 2.0;
double yc = aImage.Height / 2.0;
Boolean xpos, ypos;
//Move through the pixels in the corrected image;
//set to corresponding pixels in distorted image
for (int i = 0; i < correctedImage.Width; i++)
{
for (int j = 0; j < correctedImage.Height; j++)
{
//which quadrant are we in?
xpos = i > xc;
ypos = j > yc;
//Find the distance from the center
double xdif = i-xc;
double ydif = j-yc;
//The distance squared
double Rusquare = xdif * xdif + ydif * ydif;
//the angle from the center
double theta = Math.Atan2(ydif, xdif);
//find index for lookup table
int index = (int)(Math.Sqrt(Rusquare) / aFocalLinPixels * 1000);
if (index >= mFELimit) index = mFELimit - 1;
//calculated Rdistorted
double Rd = aFocalLinPixels * (double)mFisheyeCorrect[index]
/mScaleFESize;
//calculate x and y distances
double xdelta = Math.Abs(Rd*Math.Cos(theta));
double ydelta = Math.Abs(Rd * Math.Sin(theta));
//convert to pixel coordinates
int xd = (int)(xc + (xpos ? xdelta : -xdelta));
int yd = (int)(yc + (ypos ? ydelta : -ydelta));
xd = Math.Max(0, Math.Min(xd, aImage.Width-1));
yd = Math.Max(0, Math.Min(yd, aImage.Height-1));
//set the corrected pixel value from the distorted image
correctedImage.SetPixel(i, j, aImage.GetPixel(xd, yd));
}
}
return correctedImage;
}
}
I found this pdf file and I have proved that the maths are correct (except for the line vd = *xd**fv+v0 which should say vd = **yd**+fv+v0).
http://perception.inrialpes.fr/CAVA_Dataset/Site/files/Calibration_OpenCV.pdf
It does not use all of the latest co-efficients that OpenCV has available but I am sure that it could be adapted fairly easily.
double k1 = cameraIntrinsic.distortion[0];
double k2 = cameraIntrinsic.distortion[1];
double p1 = cameraIntrinsic.distortion[2];
double p2 = cameraIntrinsic.distortion[3];
double k3 = cameraIntrinsic.distortion[4];
double fu = cameraIntrinsic.focalLength[0];
double fv = cameraIntrinsic.focalLength[1];
double u0 = cameraIntrinsic.principalPoint[0];
double v0 = cameraIntrinsic.principalPoint[1];
double u, v;
u = thisPoint->x; // the undistorted point
v = thisPoint->y;
double x = ( u - u0 )/fu;
double y = ( v - v0 )/fv;
double r2 = (x*x) + (y*y);
double r4 = r2*r2;
double cDist = 1 + (k1*r2) + (k2*r4);
double xr = x*cDist;
double yr = y*cDist;
double a1 = 2*x*y;
double a2 = r2 + (2*(x*x));
double a3 = r2 + (2*(y*y));
double dx = (a1*p1) + (a2*p2);
double dy = (a3*p1) + (a1*p2);
double xd = xr + dx;
double yd = yr + dy;
double ud = (xd*fu) + u0;
double vd = (yd*fv) + v0;
thisPoint->x = ud; // the distorted point
thisPoint->y = vd;
This can be solved as an optimization problem. Simply draw on curves in images that are supposed to be straight lines. Store the contour points for each of those curves. Now we can solve the fish eye matrix as a minimization problem. Minimize the curve in points and that will give us a fisheye matrix. It works.
It can be done manually by adjusting the fish eye matrix using trackbars! Here is a fish eye GUI code using OpenCV for manual calibration.
With a sine input, I tried to modify it's frequency cutting some lower frequencies in the spectrum, shifting the main frequency towards zero. As the signal is not fftshifted I tried to do that by eliminating some samples at the begin and at the end of the fft vector:
interval = 1;
samplingFrequency = 44100;
signalFrequency = 440;
sampleDuration = 1 / samplingFrequency;
timespan = 1 : sampleDuration : (1 + interval);
original = sin(2 * pi * signalFrequency * timespan);
fourierTransform = fft(original);
frequencyCut = 10; %% Hertz
frequencyCut = floor(frequencyCut * (length(pattern) / samplingFrequency) / 4); %% Samples
maxFrequency = length(fourierTransform) - (2 * frequencyCut);
signal = ifft(fourierTransform(frequencyCut + 1:maxFrequency), 'symmetric');
But it didn't work as expected. I also tried to remove the center part of the spectrum, but it wielded a higher frequency sine wave too.
How to make it right?
#las3rjock:
its more like downsampling the signal itself, not the FFT..
Take a look at downsample.
Or you could create a timeseries object, and resample it using the resample method.
EDIT:
a similar example :)
% generate a signal
Fs = 200;
f = 5;
t = 0:1/Fs:1-1/Fs;
y = sin(2*pi * f * t) + sin(2*pi * 2*f * t) + 0.3*randn(size(t));
% downsample
n = 2;
yy = downsample([t' y'], n);
% plot
subplot(211), plot(t,y), axis([0 1 -2 2])
subplot(212), plot(yy(:,1), yy(:,2)), axis([0 1 -2 2])
A crude way to downsample your spectrum by a factor of n would be
% downsample by a factor of 2
n = 2; % downsampling factor
newSpectrum = fourierTransform(1:n:end);
For this to be a lower-frequency signal on your original time axis, you will need to zero-pad this vector up to the original length on both the positive and negative ends. This will be made much simpler using fftshift:
pad = length(fourierTransform);
fourierTransform = [zeros(1,pad/4) fftshift(newSpectrum) zeros(1,pad/4)];
To recover the downshifted signal, you fftshift back before applying the inverse transform:
signal = ifft(fftshift(fourierTransform));
EDIT: Here is a complete script which generates a plot comparing the original and downshifted signal:
% generate original signal
interval = 1;
samplingFrequency = 44100;
signalFrequency = 440;
sampleDuration = 1 / samplingFrequency;
timespan = 1 : sampleDuration : (1 + interval);
original = sin(2 * pi * signalFrequency * timespan);
% plot original signal
subplot(211)
plot(timespan(1:1000),original(1:1000))
title('Original signal')
fourierTransform = fft(original)/length(original);
% downsample spectrum by a factor of 2
n = 2; % downsampling factor
newSpectrum = fourierTransform(1:n:end);
% zero-pad the positive and negative ends of the spectrum
pad = floor(length(fourierTransform)/4);
fourierTransform = [zeros(1,pad) fftshift(newSpectrum) zeros(1,pad)];
% inverse transform
signal = ifft(length(original)*fftshift(fourierTransform),'symmetric');
% plot the downshifted signal
subplot(212)
plot(timespan(1:1000),signal(1:1000))
title('Shifted signal')
Plot of original and downshifted signals http://img5.imageshack.us/img5/5426/downshift.png