image processing with parallel computing gives incorrect results

i need to read all images in a folder in groups of 3 and process them for that i am using a broadcast variable to store the names of all the images, i want to do it by using parallel computing but results are different from the results i get by using simple for loop. Here is the example code, any suggestion how can i get correct results?

inDir ='./abc/';
Imgs = dir([inDir '*.jpg']);

parfor indx=2:length(Imgs)-1
    im1=imread([inDir Imgs(indx-1).name]);
    im2=imread([inDir Imgs(indx).name]);
    im3=imread([inDir Imgs(indx+1).name]);
    % resultimg=someprocessing using im1,im2 and im3
    % write resultimg
end

Answer 1

[Ahmet Cecen]

I am not very knowledgeable with the implementation of ParFor, but I would wager a guess the problem is your indexing: specifically the presence of indx-1 and indx+1.

Try to see if creating three lists of images and indexing them all with indx fixes the problem. Something like:

Imgs = dir([inDir '*.jpg']);
ImgsBefore = circshift(Imgs,1);
ImgsAfter = circshift(Imgs,-1)

parfor indx=2:length(Imgs)-1
  im1=imread([inDir ImgsBefore (indx).name]);
  im2=imread([inDir Imgs(indx).name]);
  im3=imread([inDir ImgsAfter (indx).name]);
  % resultimg=someprocessing using im1,im2 and im3
  % write resultimg
end

As I mentioned, this is simply a guess and might not work. Just come back and let it known if it doesn't.

[Maria Kanwal]

results are different than before but still do not match for loop results

[Ahmet Cecen]

I have tried a basic operation of this form (a triple convolution) and I get consistent results with for and parfor. I have no idea what else can be the problem.

Answer 2

[Walter Roberson]

inDir ='./abc/';
Imgs = dir( fullfile(inDir, '*.jpg') );

numimg = length(Imgs);

parfor indx = 1 : numimg
  im{indx} = imread( fullfile(inDir, Imgs(indx).name) );
end

parfor indx=2:length(Imgs)-1
    im1 = im{indx-1};
    im2 = im{indx};
    im3 = im{indx+1};
    % resultimg=someprocessing using im1,im2 and im3
    % write resultimg
end

If the result is not the same as before then it might have to do with the processing you are doing.

[Maria Kanwal]

having all frames in memory is not feasible for me, number of images are large that's why i wanted to use parallel computing

[Walter Roberson]

I think the differences you are observing have to do with the processing you are doing (which you did not show us.)

[Maria Kanwal]

i'm computing optical flow ( optical flow )