0
$\begingroup$

For a project I'd like to set up an easy to implement tracker. I read several papers and chose to go for the mosse (minimum output of squared error) tracker from Bolme et al.

The formula to obtain the mosse filter is (equation 5 in the paper):


H* = sum(Gi x Fi*)/(sum(Fi x Fi*) +e)

Where Gi is the fourier transform of the desired output, a 2D gaussian shaped peak centered on the target in the training image.
Fi is the fourier transform of the training image.
x denotes elementwise multiplication.
the asterisk indicates that it is the conjugate of the fourier transform.
e = 0.1, is a regularisation parameter that induces noise for enhanced noise tolerance.

To generate training images I took the first frame and applied small affine transformations to it.

But somewhere I made a mistake. Since my filter looks nothing like the one in the paper.

mine looks like this. a fish and my failed mosse filter
on the left the input image and on the right the obtained filter.

I even checked if the transformation to the training images and the desire output are faulty. However, they look fine to me.

Unfortunately I have no background in signal processing or linear algebra. So I would be really grateful if you could help me.

Below I added the Matlab code for my failed filter.


% get variables to store the numerator and denominator
sum_numerator = zeros(size(filter));
sum_denominator = zeros(size(filter));

for a = 1:numberOfTrainingImages
    Fi = fft2(image);
    Gi = fft2(desiredOutput);

    % sum_numerator is the correlation between input and desired output
    sum_numerator = sum_numerator+ ( Gi .* conj(Fi));  

    % sum_denominator is the energy spectrum of the input
    sum_denominator = sum_denominator + ( Fi .* conj(Fi));

    % Generate a training image and a according desired output
    [image, desiredOutput] = generateTrainingImage (OriginalImage);
end

% elementwise division to get the filter
H = sum_numerator ./ sum_denominator;
$\endgroup$
0
$\begingroup$

The solution to the problem was, to rearrange the quadrants of the filter. After rearranging the quadrants, my filter looks like this:

enter image description here

A lot more like the one in the paper. However, it still lacks in tracking precission. To tweak this I played arround with the 2D gaussian. And obtained the filter below. Still not optimal but it's an improvement.

enter image description here

This is the formula for the 2D gaussian $$ g_{i}(x,y) = e^{-\frac{(x-x_{i})^2+(y-y_{i})^2}{\sigma ^{2}}} $$

In the paper I mentioned in the question, Bolme et al. used $\sigma = 2$. Instead of this I obtained $\sigma$ now dependent on the size of my template

$$ \sigma = \sqrt{{imgSize_{x} \cdot {imgSize_{y}}}} \cdot \frac{1}{16} $$

like used by Galoogahi et al. (2015). Correlation Filters with Limited Boundaries

$\endgroup$
  • $\begingroup$ Although it has been some time, can I please ask you to accept this answer? That would close the question and stop it from circulating on the board. $\endgroup$ – A_A Jun 30 '18 at 9:35

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.