Optical flow evaluation demo

Computes flow field between two images using various methods and display it (deepflow, simpleflow, sparsetodenseflow, Farneback, TV-L1).

Sources:

Contents

Input images

a pair of 8-bit color images

im1 = imread(fullfile(mexopencv.root(),'test','RubberWhale1.png'));
im2 = imread(fullfile(mexopencv.root(),'test','RubberWhale2.png'));
assert(isequal(size(im1), size(im2)), ...
    'Dimension mismatch between input images');
if ~mexopencv.isOctave() && mexopencv.require('images')
    %HACK: IMSHOWPAIR not implemented in Octave
    imshowpair(im1, im2);
end

Compare the different methods

algorithms = {'farneback', 'simpleflow', 'tvl1', 'deepflow', ...
    'sparsetodenseflow', 'pcaflow', ...
    'DISflow_ultrafast', 'DISflow_fast', 'DISflow_medium', 'variational'};
for i=1:numel(algorithms)
    % prepare images
    if any(strcmp(algorithms{i}, {'farneback', 'tvl1', 'deepflow', ...
            'DISflow_ultrafast', 'DISflow_fast', 'DISflow_medium', ...
            'variational'})) && size(im1,3)==3
        % 1-channel images are expected
        img1 = cv.cvtColor(im1, 'RGB2GRAY');
        img2 = cv.cvtColor(im2, 'RGB2GRAY');
    elseif strcmp(algorithms{i}, 'simpleflow') && size(im1,3)==1
        % 3-channel images expected
        img1 = cv.cvtColor(im1, 'GRAY2RGB');
        img2 = cv.cvtColor(im2, 'GRAY2RGB');
    else
        % sparsetodenseflow/pcaflow handle both 1- or 3-channels
        img1 = im1;
        img2 = im2;
    end

    % compute flow field between img1 and img2 using current method
    tic
    switch lower(algorithms{i})
        case 'farneback'
            %{
            obj = cv.FarnebackOpticalFlow();
            flow = obj.calc(img1, img2);
            %}
            flow = cv.calcOpticalFlowFarneback(img1, img2);
        case 'simpleflow'
            flow = cv.calcOpticalFlowSF(img1, img2);
        case 'deepflow'
            flow = cv.calcOpticalFlowDF(img1, img2);
        case 'sparsetodenseflow'
            flow = cv.calcOpticalFlowSparseToDense(img1, img2);
        case 'tvl1'
            obj = cv.DualTVL1OpticalFlow();
            flow = obj.calc(img1, img2);
        case 'pcaflow'
            %obj = cv.OpticalFlowPCAFlow(prior);  % path to a prior file for PCAFlow
            obj = cv.OpticalFlowPCAFlow();
            flow = obj.calc(img1, img2);
        case 'disflow_ultrafast'
            obj = cv.DISOpticalFlow('Preset','UltraFast');
            flow = obj.calc(img1, img2);
        case 'disflow_fast'
            obj = cv.DISOpticalFlow('Preset','Fast');
            flow = obj.calc(img1, img2);
        case 'disflow_medium'
            obj = cv.DISOpticalFlow('Preset','Medium');
            flow = obj.calc(img1, img2);
        case 'variational'
            obj = cv.VariationalRefinement();
            flow = obj.calc(img1, img2);
    end
    fprintf('%18s: ', algorithms{i});
    toc

    % display the flow
    if true
        [mag, ang] = cv.cartToPolar(flow(:,:,1), flow(:,:,2), 'Degrees',true);
    else
        [ang, mag] = cart2pol(flow(:,:,1), flow(:,:,2));
        if mexopencv.isOctave()
            %HACK: RAD2DEG not implemented in Octave
            ang = (ang + pi) * (180 / pi);
        else
            ang = rad2deg(ang + pi);
        end
    end
    mag = cv.normalize(mag, 'Alpha',0, 'Beta',1, 'NormType','MinMax');
    hsv = cat(3, ang, ones(size(ang),class(ang)), mag); % H=[0,360], S,V=[0,1]
    rgb = cv.cvtColor(hsv, 'HSV2RGB');                  % R,G,B=[0,1]
    figure, imshow(rgb)
    title(sprintf('Computed flow: %s',algorithms{i}), 'Interpreter','none')
end

         farneback: Elapsed time is 0.201330 seconds.
        simpleflow: Elapsed time is 1.529157 seconds.
              tvl1: Elapsed time is 1.269737 seconds.
          deepflow: Elapsed time is 0.849141 seconds.
 sparsetodenseflow: Elapsed time is 0.341761 seconds.
           pcaflow: Elapsed time is 0.625318 seconds.
 DISflow_ultrafast: Elapsed time is 0.124396 seconds.
      DISflow_fast: Elapsed time is 0.020869 seconds.
    DISflow_medium: Elapsed time is 0.046709 seconds.
       variational: Elapsed time is 0.149174 seconds.