Required to turn an image into N triangles with Delaunay triangulation. One color for each triangle, and colors can be repeated.
The loss function is given by the square of the difference in the color of each pixel.
So how to optimize the color and the vertices of triangles?
A recursive splitting procedure outline:
Terminate the recursion if N < 2
Split the given area A in two triangles A1 and A2 in such a way that the
sum of standard deviations of the pixel colors is cut in halves.
Assign N/2 colors to A1 and N - N/2 colors to A2.
Recursively split A1 and A2.
The resulting net of N triangles is colored to minimize the loss function:
For every triangle the color chosen is the average color of the pixels within that triangle.
It might be worthwhile to conduct a survey of existing literature on the topic. A first search engine hit returned Fractal image compression based on Delaunay triangulation and vector quantization
Related
I have a tensor of shape (N, 3, 3) where N is the number of triangles. Is there a way to calculate the area of a triangle from a 3x3 matrix of coordinates so I can batch along the N dimension?
Most methods for vectorizing area computations rely on subtracting one of the vertices from the other two and taking the cross product which I don't think is possible to do with batching.
I have a 2D matrix of 2-tuples representing x,y coordinates which I plot using plt.scatter(), resulting in a square grid of points. I'd like to color all points in different colors in such a way so nearby points (Euclidean or Manhattan distance, doesn't really matter) are colored in similar colors, while taking advantage of a wide as possible range of colors for the whole grid (so points in general are as distinguishable as possible). How can this be achieved?
I'm needing to implement a Minkowski sum function that can return the Minkowski sum of either 2 circles, 2 convex polygons or a circle and a convex polygon. I found this thread that explained how to do this for convex polygons, but I'm not sure how to do this for a circle and polygon. Also, how would I even represent the answer?! I'd like the algorithm to run in O(n) time but beggars can't be choosers.
Circle is trivial -- just add the center points, and add the radii. Circle + ConvexPoly is nearly as simple: move each segment perpendicularly outward by the circle radius, and connect adjacent segments with circular arcs centered at the original poly vertices. Translate the whole by the circle center point.
As for how you represent the answer: Well, it depends on what you want to do with it. You could convert it to a NURBS if you just want to draw it with a vector drawing library. You could approximate the circular arcs with polylines if you just want a polygonal approximation. Or you might store it as is -- "this polygon, expanded by such-and-such a radius". That would be the best choice for things like raycasting, for instance. Or as a compromise, you could connect adjacent segments linearly instead of with circular arcs, and store it as the union of the (new) convex polygon and a list of circles at the vertices.
Oh, about ConvexPoly + ConvexPoly. That's the trickiest one, but still straightforward. The basic idea is that you take the list of segment vectors for each polygon (starting from some particular extremal point, like the point on each poly with the lowest X coordinate), then merge the two lists together, keeping it sorted by angle. Sum the two points you started with, then apply each vector from the merged vector list to produce the other points.
I have a number of 2D (possibly intersecting) polygons which I rendered using OpenGL ES on the screen. All the polygons are completely contained within the screen. What is the most timely way to find the percentage area of the union of these polygons to the total screen area? Timeliness is required as I have a requirement for the coverage area to be immediately updated whenever a polygon is shifted.
Currently, I am representing each polygon as a 2D array of booleans. Using a point-in-polygon function (from a geometry package), I sample each point (x,y) on the screen to check if it belongs to the polygon, and set polygon[x][y] = true if so, false otherwise.
After doing that to all the polygons in the screen, I loop through all the screen pixels again, and check through each polygon array, counting that pixel as "covered" if any polygon has its polygon[x][y] value set to true.
This works, but the performance is not ideal as the number of polygons increases. Are there any better ways to do this, using open-source libraries if possible? I thought of:
(1) Unioning the polygons to get one or more non-overlapping polygons. Then compute the area of each polygon using the standard area-of-polygon formula. Then sum them up. Not sure how to get this to work?
(2) Using OpenGL somehow. Imagine that I am rendering all these polygons with a single color. Is it possible to count the number of pixels on the screen buffer with that certain color? This would really sound like a nice solution.
Any efficient means for doing this?
If you know background color and all polygons have other colors, you can read all pixels from framebuffer glReadPixels() and simply count all pixels that have color different than background.
If first condition is not met you may consider creating custom framebuffer and render all polygons with the same color (For example (0.0, 0.0, 0.0) for backgruond and (1.0, 0.0, 0.0) for polygons). Next, read resulting framebuffer and calculate mean of red color across the whole screen.
If you want to get non-overlapping polygons, you can run a line intersection algorithm. A simple variant is the Bentley–Ottmann algorithm, but even faster algorithms of O(n log n + k) (with n vertices and k crossings) are possible.
Given a line intersection, you can unify two polygons by constructing a vertex connecting both polygons on the intersection point. Then you follow the vertices of one of the polygons inside of the other polygon (you can determine the direction you have to go in using your point-in-polygon function), and remove all vertices and edges until you reach the outside of the polygon. There you repair the polygon by creating a new vertex on the second intersection of the two polygons.
Unless I'm mistaken, this can run in O(n log n + k * p) time where p is the maximum overlap of the polygons.
After unification of the polygons you can use an ordinary area function to calculate the exact area of the polygons.
I think that attempt to calculate area of polygons with number of pixels is too complicated and sometimes inaccurate. You can see something similar in stackoverflow answer about calculation the area covered by a polygon and if you construct regular polygons see area of a regular polygon ,
I am trying to plot a matrix in Gnuplot as I would using imshow in Matplotlib. That means I just want to plot the actual matrix values, not the interpolation between values. I have been able to do this by trying
splot "file.dat" u 1:2:3 ps 5 pt 5 palette
This way we are telling the program to use columns 1,2 and 3 in the file, use squares of size 5 and space the points with very narrow gaps. However the points in my dataset are not evenly spaced and hence I get discontinuities.
Anyone a method of plotting matrix values in gnuplot regardless of not evenly spaced in Xa and y axes?
Gnuplot doesn't need to have evenly space X and Y axes. ( see another one of my answers: https://stackoverflow.com/a/10690041/748858 ). I frequently deal with grids that look like x[i] = f_x(i) and y[j] = f_y(j). This is quite trivial to plot, the datafile just looks like:
#datafile.dat
x1 y1 z11
x1 y2 z12
...
x1 yN z1N
#<--- blank line (leave these comments out of your datafile ;)
x2 y1 z21
x2 y2 z22
...
x2 yN z2N
#<--- blank line
...
...
#<--- blank line
xN y1 zN1
...
xN yN zNN
(note the blank lines)
A datafile like that can be plotted as:
set view map
splot "datafile.dat" u 1:2:3 w pm3d
the option set pm3d corners2color can be used to fine tune which corner you want to color the rectangle created.
Also note that you could make essentially the same plot doing this:
set view map
plot "datafile.dat" u 1:2:3 w image
Although I don't use this one myself, so it might fail with a non-equally spaced rectangular grid (you'll need to try it).
Response to your comment
Yes, pm3d does generate (M-1)x(N-1) quadrilaterals as you've alluded to in your comment -- It takes the 4 corners and (by default) averages their value to assign a color. You seem to dislike this -- although (in most cases) I doubt you'd be able to tell a difference in the plot for reasonably large M and N (larger than 20). So, before we go on, you may want to ask yourself if it is really necessary to plot EVERY POINT.
That being said, with a little work, gnuplot can still do what you want. The solution is to specify that a particular corner is to be used to assign the color to the entire quadrilateral.
#specify that the first corner should be used for coloring the quadrilateral
set pm3d corners2color c1 #could also be c2,c3, or c4.
Then simply append the last row and last column of your matrix to plot it twice (making up an extra gridpoint to accommodate the larger dataset. You're not quite there yet, you still need to shift your grid values by half a cell so that your quadrilaterals are centered on the point in question -- which way you shift the cells depends on your choice of corner (c1,c2,c3,c4) -- You'll need to play around with it to figure out which one you want.
Note that the problem here isn't gnuplot. It's that there isn't enough information in the datafile to construct an MxN surface given MxN triples. At each point, you need to know it's position (x,y) it's value (z) and also the size of the quadrilateral to be draw there -- which is more information than you've packed into the file. Of course, you can guess the size in the interior points (just meet halfway), but there's no guessing on the exterior points. but why not just use the size of the next interior point?. That's a good question, and it would (typically) work well for rectangular grids, but that is only a special case (although a common one) -- which would (likely) fail miserably for many other grids. The point is that gnuplot decided that averaging the corners is typically "close enough", but then gives you the option to change it.
See the explanation for the input data here. You may have to change your data file's format accordingly.