I am doing some lane detection as a fun project and I am trying to create a cubic spline to represent the lane. However, when I use the scipy CubicSpline function I get some absurdly large values.
Here is my code:
from scipy import interpolate
from scipy.interpolate import CubicSpline
from scipy.interpolate import InterpolatedUnivariateSpline
from scipy.interpolate import interp1d
rows = img_size[0] # width
height = left_lane.shape[0]
y_values = [0, height/2, height]
plt.figure()
plt.imshow(left_lane, cmap='gray')
splines = []
particle = particles[0]
cx = [particle[0], particle[1], particle[2]]
cy = [y_values[0], y_values[1], y_values[2]]
points = zip(cx, cy)
points = sorted(points, key=lambda point: point[0])
x1, y1 = zip(*points)
x1 = np.asarray(x1)
y1 = np.asarray(y1)
s = CubicSpline(x1, y1)
new_x = np.arange(min(x1), max(x1), 0.1)
new_y = s(new_x)
plt.plot(new_x, new_y)
plt.show()
Here is the output:
Here is the original image with the control points plotted:
I don't understand why the spline algorithm is giving me such large values for such a simple spline. So, what is the issue here? Bad values? Incorrect usage of cubic spline?
Thank you for the help!
The issue was that the spline is going vertical in this image which means that the x value isn't in always increasing order (e.g. 2nd control > 1st control point). So, when I sorted my lists something like this could happen (2, 1, 3). Then, scipy has to use some huge coefficients to fit this spline.
The fix was to flip the axis so that y is now the x axis. Now the sorting doesn't cause the any weird ordering of the control points.
cx = [particle[0], particle[1], particle[2]]
cy = [y_values[0], y_values[1], y_values[2]]
# Sort particles in increasing x order
points = zip(cx, cy)
points = sorted(points, key=lambda point: point[1])
x1, y1 = zip(*points)
x1 = np.asarray(x1)
y1 = np.asarray(y1)
s = CubicSpline(y1, x1)
new_x = np.arange(min(y1), max(y1), 0.1)
new_y = s(new_x)
plt.plot(new_y, new_x)
plt.plot(cx, cy, '.')
Related
I am currently trying to produce a 2D heat transfer analysis on an alloy which has a shape of a semicircle, (and then introduce a heat source on this shape). I found myself unable to produce a mesh for that specific shape, the code below shows what I have done up until now, if anyone could possibly point me in the right direction it would be greatly appreciated!
initially I set up the grid parameters as follows:
%-------------------GRID SETTINGS----------------------%
Sx = 10; %physical width
Sy = 10; %physical height
Nx = 100; %number of points x-axis
Ny = round(Nx*Sy/Sx); %number of points y-axis
Nt = 500;
%-------------------GRID RESOLUTION---------------------%
dx = Sx/(Nx);
dy = Sy/(Ny);
and worked on creating the shape of an enclosed semicircle as below:
%---------------------SHAPE CREATION--------------------%
r1 = 0.07; %radius 1, m
r2 = 5; %radius 2, m
phi = 0:0.01:pi;
x = r2*cos(phi); x = [x,x(1)]; %x coordinates
y = r1*sin(phi); y = [y,y(1)]; %y coordinates
[X, Y] = meshgrid(x,y);
I'd use polar coordinates for your mesh and then visualize the results in cartesian coordinates. Check out the example below, hope this gives you a starting point (not really sure why you applied two different radius (r1, r2) to your x,y coordinates - so I assumed you have a minimal and maximal radius).
Nx = 100;
r1 = 1; %radius 1, m
r2 = 5; %radius 2, m
phi = 0:0.01:pi;
% generate mesh grid in polar coordinates
radius = linspace(r1, r2, Nx/2);
[Radius, Phi] = meshgrid(radius, phi);
heat = Radius.^2; % some "fake" heat data
% plot results (in cartesian coordinates)
mesh(Radius.*cos(Phi), Radius.*sin(Phi), heat)
axis equal
% top view
view([-0.2 90.0])
How to distribute the points to be like Fig.A
This matlab code for Fig. B :
N = 30; % number of points
r = 0.5; % r = radius
d = 50; % dimension
C_point = 0; % center point
figure, clf
C = ones(1, d) * C_point;
C_rep = repmat( C,N,1);
X = randn(N,d);
s2 = sum(X.^2,2) ;
radius = r * (rand(N,1).^(1/d));
X = X.*repmat(radius./sqrt(s2),1,d) + C_rep;
%% Plot 2D
t = linspace(0, 2*pi, 100);
x = r*cos(t) + C(1);
y = r*sin(t) + C(2);
plot(x,y,'b')
hold on
plot(C(1),C(2),'b.', 'MarkerSize', 10) % center point
hold on
plot(X(:,1), X(:,2),'r.','markersize',10);
axis equal;rotate3d off; rotate3d on;drawnow;shg;
hold on
ax = axis;
Source of the code
What I should change to be like fig. A
The OP's code computes points uniformly distributed within a d-dimensional box, projects those onto a d-dimensional sphere, then samples the radius to move them inside the d-dimensional ball. This is perfect except that the points inside the box, when projected onto the sphere, do not form a uniform distribution on that sphere. If instead you find random points distributed in a Gaussian distribution, you are guaranteed uniform angle distribution.
First compute points with a Gaussian distribution in d dimensions (I do all here with minimal changes to the OP's code):
N = 1000; % number of points
r = 0.5; % r = radius
d = 3; % dimension
C_point = 0; % center point
C = ones(1,d) * C_point;
C_rep = repmat(C,N,1);
X = randn(N,d);
Note that I use randn, not rand. randn creates a Gaussian distribution.
Next we normalize the vectors so the points move to the sphere:
nX = sqrt(sum(X.^2,2));
X = X./repmat(nX,1,d);
These points are uniformly distributed, which you can verify by scatter3(X(:,1),X(:,2),X(:,3)); axis equal and turning the display around (a 2D rendering doesn't do it justice). This is the reason I set d=3 above, and N=1000. I wanted to be able to plot the points and see lots of them.
Next we compute, as you already did, a random distance to the origin, and correct it for the dimensionality:
radius = r * (rand(N,1).^(1/d));
X = X.*repmat(radius,1,d) + C_rep;
X now is distributed uniformly in the ball. Again, scatter3(X(:,1),X(:,2),X(:,3)); axis equal shows this.
However, if you set d=50 and then plot only two dimensions of your data, you will not see the data filling the circle. And you will not see a uniform distribution either. This is because you are projecting a 50-D ball onto 2 dimensions, this simply does not work. You either have to trust the math, or you have to slice the data:
figure, hold on
t = linspace(0, 2*pi, 100);
x = r*cos(t) + C(1);
y = r*sin(t) + C(2);
plot(x,y,'b')
plot(C(1),C(2),'b.', 'MarkerSize', 10) % center point
axis equal
I = all(abs(X(:,3:d))<0.1,2);
plot(X(I,1), X(I,2),'r.','markersize',10);
The I there indexes points that are close to the origin in dimensions perpendicular to the first two shown. Again, with d=50 you will have very few points there, so you will need to set N very large! To see the same density of points as in the case above, for every dimension you add, you need to multiply N by 10. So for d=5 you'd have N=1000*10*10=1e5, and for d=50 you'd need N=1e50. That is totally impossible to compute, of course.
I have polar coordinates, radius 0.05 <= r <= 1 and 0 ≤ θ ≤ 2π. The radius r is 50 values between 0.05 to 1, and polar angle θ is 24 values between 0 to 2π.
How do I interpolate r = 0.075 and theta = pi/8?
I dunno what you have tried, but interp2 works just as well on polar data as it does on Cartesian. Here is some evidence:
% Coordinates
r = linspace(0.05, 1, 50);
t = linspace(0, 2*pi, 24);
% Some synthetic data
z = sort(rand(50, 24));
% Values of interest
ri = 0.075;
ti = pi/8;
% Manually interpolate
rp = find(ri <= r, 1, 'first');
rm = find(ri >= r, 1, 'last');
tp = find(ti <= t, 1, 'first');
tm = find(ti >= t, 1, 'last');
drdt = (r(rp) - r(rm)) * (t(tp) - t(tm));
dr = [r(rp)-ri ri-r(rm)];
dt = [t(tp)-ti ti-t(tm)];
fZ = [z(rm, tm) z(rm, tp)
z(rp, tm) z(rp, tp)];
ZI_manual = (dr * fZ * dt.') / drdt
% Interpolate with MATLAB
ZI_MATLAB = interp2(r, t, z', ri, ti, 'linear')
Result:
ZI_manual =
2.737907208525297e-002
ZI_MATLAB =
2.737907208525298e-002
Based on comments you have the following information
%the test point
ri=0.53224;
ti = pi/8;
%formula fo generation of Z
g=9.81
z0=#(r)0.01*(g^2)*((2*pi)^-4)*(r.^-5).*exp(-1.25*(r/0.3).^-4);
D=#(t)(2/pi)*cos(t).^2;
z2=#(r,t)z0(r).*D(t) ;
%range of vlaues of r and theta
r=[0.05,0.071175,0.10132,0.14422,0.2053, 0.29225,0.41602,0.5922,0.84299,1.2];
t=[0,0.62832,1.2566,1.885, 2.5133,3.1416,3.7699,4.3982,5.0265,5.6549,6.2832];
and you want interplation of the test point.
When you sample some data to use them for interpolation you should consider how to sample data according to your requirements.
So when you are sampling a regular grid of polar coordinates ,those coordinates when converted to rectangular will form a circular shape that
most of the points are concentrated in the center of the cricle and when we move from the center to outer regions distance between the points increased.
%regular grid generated for r and t
[THETA R] = meshgrid(t ,r);
% Z for polar grid
Z=z2(R,THETA);
%convert coordinate from polar to cartesian(rectangular):
[X, Y] = pol2cart (THETA, R);
%plot points
plot(X, Y, 'k.');
axis equal
So when you use those point for interpolation the accuracy of the interpolation is greater in the center and lower in the outer regions where the distance between points increased.
In the other word with this sampling method you place more importance on the center region related to outer ones.
To increase accuracy density of grid points (r and theta) should be increased so if length of r and theta is 11 you can create r and theta with size 20 to increase accuracy.
In the other hand if you create a regular grid in rectangular coordinates an equal importance is given to each region . So accuracy of the interpolation will be the same in all regions.
For it first you create a regular grid in the polar coordinates then convert the grid to rectangular coordinates so you can calculate the extents (min max) of the sampling points in the rectangular coordinates. Based on this extents you can create a regular grid in the rectangular coordinates
Regular grid of rectangular coordinates then converted to polar coordinated to get z for grid points using z2 formula.
%get the extent of points
extentX = [min(X(:)) max(X(:))];
extentY = [min(Y(:)) max(Y(:))];
%sample 100 points(or more or less) inside a region specified be the extents
X_samples = linspace(extentX(1),extentX(2),100);
Y_samples = linspace(extentY(1),extentY(2),100);
%create regular grid in rectangular coordinates
[XX YY] = meshgrid(X_samples, Y_samples);
[TT RR] = cart2pol(XX,YY);
Z_rect = z2(RR,TT);
For interpolation of a test point say [ri ti] first it converted to rectangular then using XX ,YY value of z is interpolated
[xi yi] = pol2cart (ti, ri);
z=interp2(XX,YY,Z_rect,xi,yi);
If you have no choice to change how you sample the data and only have a grid of polar points as discussed with #RodyOldenhuis you can do the following:
Interpolate polar coordinates with interp2 (interpolation for gridded data)
this approach is straightforward but has the shortcoming that r and theta are not of the same scale and this may affect the accuracy of the interpolation.
z = interp2(THETA, R, Z, ti, ri)
convert polar coordinates to rectangular and then apply an interpolation method that is for scattered data.
this approach requires more computations but result of it is more reliable.
MATLAB has griddata function that given scattered points first generates a triangulation of points and then creates a regular grid on top of the triangles and interpolates values of grid points.
So if you want to interpolate value of point [ri ti] you should then apply a second interpolation to get value of the point from the interpolated grid.
With the help of some information from spatialanalysisonline and Wikipedia linear interpolation based on triangulation calculated this way (tested in Octave. In newer versions of MATLAB use of triangulation and pointLocation recommended instead of delaunay and tsearch ):
ri=0.53224;
ti = pi/8;
[THETA R] = meshgrid(t ,r);
[X, Y] = pol2cart (THETA, R);
[xi yi] = pol2cart (ti, ri);
%generate triangulation
tri = delaunay (X, Y);
%find the triangle that contains the test point
idx = tsearch (X, Y, tri, xi, yi);
pts= tri(idx,:);
%create a matrix that repesents equation of a plane (triangle) given its 3 points
m=[X(pts);Y(pts);Z(pts);ones(1,3)].';
%calculate z based on det(m)=0;
z= (-xi*det(m(:,2:end)) + yi*det([m(:,1) m(:,3:end)]) + det(m(:,1:end-1)))/det([m(:,1:2) m(:,end)]);
More refinement:
Since it is known that the search point is surrounded by 4 points we can use only those point for triangulation. these points form a trapezoid. Each diagonal of trapezoid forms two triangles so using vertices of the trapezoid we can form 4 triangles, also a point inside a trapezoid can lie in at least 2 triangles.
the previous method based on triangulation only uses information from one triangle but here z of the test point can be interpolated two times from data of two triangles and the calculated z values can be averaged to get a better approximation.
%find 4 points surrounding the test point
ft= find(t<=ti,1,'last');
fr= find(cos(abs(diff(t(ft+(0:1))))/2) .* r < ri,1,'last');
[T4 R4] = meshgrid(t(ft+(0:1)), r(fr+(0:1)));
[X4, Y4] = pol2cart (T4, R4);
Z4 = Z(fr+(0:1),ft+(0:1));
%form 4 triangles
tri2= nchoosek(1:4,3);
%empty vector of z values that will be interpolated from 4 triangles
zv = NaN(4,1);
for h = 1:4
pts = tri2(h,:);
% test if the point lies in the triangle
if ~isnan(tsearch(X4(:),Y4(:),pts,xi,yi))
m=[X4(pts) ;Y4(pts) ;Z4(pts); [1 1 1]].';
zv(h)= (-xi*det(m(:,2:end)) + yi*det([m(:,1) m(:,3:end)]) + det(m(:,1:end-1)))/det([m(:,1:2) m(:,end)]);
end
end
z= mean(zv(~isnan(zv)))
Result:
True z:
(0.0069246)
Linear Interpolation of (Gridded) Polar Coordinates :
(0.0085741)
Linear Interpolation with Triangulation of Rectangular Coordinates:
(0.0073774 or 0.0060992) based on triangulation
Linear Interpolation with Triangulation of Rectangular Coordinates(average):
(0.0067383)
Conclusion:
Result of interpolation related to structure of original data and the sampling method. If the sampling method matches pattern of original data result of interpolation is more accurate, so in cases that grid points of polar coordinates follow pattern of data result of interpolation of regular polar coordinate can be more reliable. But if regular polar coordinates do not match the structure of data or structure of data is such as an irregular terrain, method of interpolation based on triangulation can better represent the data.
please check this example, i used two for loops, inside for loop i used condition statement, if u comment this condition statement and run the program, u'll get correct answer, after u uncomment this condition statement and run the program, u'll get wrong answer. please check it.
% Coordinates
r = linspace(0.05, 1, 10);
t = linspace(0, 2*pi, 8);
% Some synthetic data
%z = sort(rand(50, 24));
z=zeros();
for i=1:10
for j=1:8
if r(i)<0.5||r(i)>1
z(i,j)=0;
else
z(i,j) = r(i).^3'*cos(t(j)/2);
end
end
end
% Values of interest
ri = 0.55;
ti = pi/8;
% Manually interpolate
rp = find(ri <= r, 1, 'first');
rm = find(ri >= r, 1, 'last');
tp = find(ti <= t, 1, 'first');
tm = find(ti >= t, 1, 'last');
drdt = (r(rp) - r(rm)) * (t(tp) - t(tm));
dr = [r(rp)-ri ri-r(rm)];
dt = [t(tp)-ti ti-t(tm)];
fZ = [z(rm, tm) z(rm, tp)
z(rp, tm) z(rp, tp)];
ZI_manual = (dr * fZ * dt.') / drdt
% Interpolate with MATLAB
ZI_MATLAB = interp2(r, t, z', ri, ti, 'linear')
Result:
z1 =
0.1632
ZI_manual =
0.1543
ZI_MATLAB =
0.1582
I need to fit data e.g. x, y, CI (where CI is confidence index of y) in Matlab.
Now, I use this code:
pf = polyfit(x, y, 2);
x1 = min(x):.1:max(x);
y1 = polyval(pf, x1);
figure
hold on
errorbar(x, y, CI, 'ko');
plot(x1, y1, 'k');
hold off
Of course, the fit comes out of some errors bars, and it's correct.
I would like obtain a fit curve closer to the points with a low confidence index, and discard the points with a high confidence index.
Thank you and bye,
Giacomo
What you are looking for are Weighted Least Squares. You can compute them with the function lscov. There is a nice example in its help page, but I'll try to make it clearer.
Let us construct a simple parabola, with a corrupted point
x = (0:0.1:1)';
y = 0.5*x.^2;
y(5) = 3*y(5);
and give some weights
w = ones(size(y));
w(5) = 0.1;
Next build the Vandermonde matrix (see here for the code) and solve the system
%// V = [x.^2 x ones(size(x))];
V = bsxfun(#power, x, 2:-1:0);
coeff = lscov(V, y, w);
The estimated coefficients, with and without the weights, are
x^2 x 1
with weights [0.4797 0.0186 -0.0004]
no weights [0.3322 0.1533 -0.0034]
Note that in your case w will have to be inverted.
If you don't like to build the Vandermonde matrix, and you have a license for the Curve Fitting Toolbox, you can use the following code
ft = fittype('poly2');
opts = fitoptions('Method', 'LinearLeastSquares');
opts.Weights = w;
fitresult = fit(x, y, ft, opts);
and you'll obtain the same result.
I have a set of 2D points (not ordered) forming a closed contour, and I would like to resample them to 14 equally spaced points. It is a contour of a kidney on an image. Any ideas?
One intuitive approach (IMO) is to create an independent variable for both x and y. Base it on arc length, and interpolate on it.
% close the contour, temporarily
xc = [x(:); x(1)];
yc = [y(:); y(1)];
% current spacing may not be equally spaced
dx = diff(xc);
dy = diff(yc);
% distances between consecutive coordiates
dS = sqrt(dx.^2+dy.^2);
dS = [0; dS]; % including start point
% arc length, going along (around) snake
d = cumsum(dS); % here is your independent variable
perim = d(end);
Now you have an independent variable and you can interpolate to create N segments:
N = 14;
ds = perim / N;
dSi = ds*(0:N).'; %' your NEW independent variable, equally spaced
dSi(end) = dSi(end)-.005; % appease interp1
xi = interp1(d,xc,dSi);
yi = interp1(d,yc,dSi);
xi(end)=[]; yi(end)=[];
Try it using imfreehand:
figure, imshow('cameraman.tif');
h = imfreehand(gca);
xy = h.getPosition; x = xy(:,1); y = xy(:,2);
% run the above solution ...
Say your contour is defined by independent vector x and dependent vector y.
You can get your resampled x vector using linspace:
new_x = linspace(min(x),max(x),14); %14 to get 14 equally spaced points
Then use interp1 to get new_y values at each new_x point:
new_y = interp1(x,y,new_x);
There are a few interpolation methods to choose from - default is linear. See interp1 help for more info.