Vectorize a triangle scan line in MATLAB/Octave

Question

I am new to Octave/MATLAB. I want to vectorize a simple scan-line fill for a triangle. Here is the while loop I want to eliminate:

 #Iterate over all the scan lines from bottom to top
 while(yCurr <= 200)

    #VARIABLES: 
    #  img - Zero matrix used as an image. Zero is black. Assume 320x200 pixels.
    #  leftIdx - Vector that contains the left-most pixel to fill in each
    #      scanline. Dim is 1x200. Each element represents x-coord of a pixel.
    #  rightIdx - Vector that contains the right-most pixel to fill in each
    #      scanline. Dim is also 1x200. Each element represents x-coord of a pixel.
    #  yCurr - The current row being scanned.    

    #Fill all the pixels in one scan line 
    img(leftIdx(yCurr) : rightIdx(yCurr), yCurr) = color;

    #Increment the row
    yCurr++;

 endwhile

angainor · Accepted Answer · 2012-09-18T22:04:02.413

Simple solution - you can change your loop directly to one arrayfun call. arrayfun is a fancy way of writing a loop - it calls user-defined function for all supplied parameters. The caveat is that anonymous functions that one usually uses with arrayfun can not assign variables. But, you can write a normal function that assigns color to img and pass it to arrayfun as parameter:

 function doAllWork(img, rightIdx, leftIdx)

 arrayfun(@fill, 1:size(img, 1));

    function fill(i)
      img(leftIdx(i):rightIdx(i), i) = color;
    end
 end

I have defined the fill function as a local function in doAllWork so that it has access to img, leftIdx, rightIdx.

Complicated solution Usually in such cases, to do an assignment in one go you want to use sub2ind to get a linear index into the matrix (in short, in C you would write sth. like j*nx+i). You then write img(ind) instead of img(indrow, indcol). The problem is that every row has a different number of non-zeros in different places.

The (complicated) idea is to create an explicit column index array Ii{row} = [leftIdx(row):rightIdx(row)] and corresponding row index array Ij{row} = [row*ones(1, lenght(Ii{row}))] for all rows. Without a loop that can be done using arrayfun. Once you have this, you can construct a linear index into your img using sub2ind on coresponding Ii/Ij entries, also called using arrayfun. The code looks like this

nrows=1:size(img, 1);
Ii=arrayfun(@(i)(leftIdx(i):rightIdx(i)), nrows, 'UniformOutput', false);
Ij=arrayfun(@(i)(i*ones(1,length(Ii{i}))), nrows, 'UniformOutput', false);
lin=arrayfun(@(i)(sub2ind(size(A), Ij{i}, Ii{i})), nrows, 'UniformOutput', false);
img([lin{:}])=color;

This approach is not of much use in your case - it is way too complicated. But it is instructive in what arrayfun can do, and the trick with sub2ind is generally very useful.

Thank your for the answer. I grok with simple solution and had considered using arrayfun. I admit I do not understand yet the complex solution yet, but I will look into sub2ind as it shows up frequently in "hints and tips" sections. The solution allows me to avoid the loop, but I have a hidden agenda of parallelizing the code. I wonder if Octave is smart enough to divide the arrayfun work over multiple cores? Answer voted up. — ahoffer, Sep 19 '12 at 16:58

Andrey Rubshtein · Answer 2 · 2012-09-19T17:12:39.203

1

I think that your original loop-based solution is the best solution. Why?

Because until you haven't done profiling you cannot know whether a vectorized solution is actually faster.
Because your code that uses for loop is clear for the reader! It contains only 4 lines of code and one can understand its purpose merely by taking a look.

The only thing that I would change is while->for

for yCurr = 1:200
    img(leftIdx(yCurr) : rightIdx(yCurr), yCurr) = color;
 end

edited Sep 19 '12 at 17:12

answered Sep 18 '12 at 20:49

Andrey Rubshtein

20,795
11
69
104

Well said. I do not know if the loop is a performance bottleneck. Performance is not my first concern. I want to learn the MATLAB programming paradigm and I have the sense that often loops can be turned into matrix ops. But not always. :-). I am up-voting this answer for its enlightened software engineering practices. – ahoffer Sep 19 '12 at 17:11

score 1 · Answer 3 · answered Sep 18 '12 at 21:39

Although not the OPs main question (i.e. how to vectorize the raster scan), a clean looking solution in MATLAB for filling in values in a triangle can come from using poly2mask to specify the area inside the triangle, using the triangle vertices, and use logical indexing to assign values.

% Example values
yCurr = 1:200;
img = zeros(320, 200);
leftIdx = ones(1,200);
rightIdx = 1:200;
color = 1;

% Define polygon vertices and get mask
x = [leftIdx(1) leftIdx(200) rightIdx(1) rightIdx(200)];
y = [1 1 200 200];
bw = poly2mask(x,y,320,200);
% Assign color values
img(bw) = color;

Since poly2mask most probably uses some underlying for, it is not a vectorized version of a rasterizing scan. However, it offers the benefit of handling more complex polygonal areas.

Great call. I did not know about poly2mask. I am selecting the answer that spoke to vectorization, but I have decided to dump my own scan line function and use poly2mask. I am up-voting your answer. Thanks. — ahoffer, Sep 19 '12 at 17:17

Vectorize a triangle scan line in MATLAB/Octave

3 Answers3