A3 common mistakes
Grader: florian@cim.mcgill.ca 


Q1

(a) The most common mistake observed here was that in global image 
    registration some submissions estimated translation vectors
    (hxi, hyi) for individual pixels, and reported (hx, hy) as the 
    average the individual estimates. 

(b) Many submissions included iterations over all pixels, which made their
    code very slow. Instead of doing double for loops, one could do something 
    like this:

    ind = sqrErr < minSqrErr;
    minSqrErr(ind) = sqrErr(ind);
    bestSigmas(ind) = sigma;

    In general, this question was well done. Some students used 
    uniform neighbourhoods of ones, while others used gaussian 
    neighbourhoods with high variance, which made the estimated blur
    field much smoother. 

(c) By far, the most common omission here was not mentioning the blur
    width formula, which relates estimated blur width at a pixel with 
    the depth of the pixel in the scene. 

    Another problem here was that some submission did not have a good 
    discussion of the limitations of trying to estimate depth from defocus.

Q2

(a) Some submissions had problems with the computation of the 
    laplacian of a function, which could be done like this:

    Ixx = conv2(conv2(Iblur, [.5 0 -.5], 'same'), [.5 0 -.5], 'same');
    Iyy = conv2(conv2(Iblur, [.5 0 -.5]', 'same'), [.5 0 -.5]', 'same');
    LoG = sigma^2 * (Ixx + Iyy);

    or like this 

    log_kernel = [0 1 0; 1 -4 1; 0 1 0];
    LoG = sigma^2 * conv2(Iblur, log_kernel, 'same');

    There were some cases, where students did not scale the LoG by sigma^2
    and as a result, their local extrema detection was scale-dependent. 

    Another very common problem was the accumulation of giant conditions
    in if statements for the checking of local extrema. For relative local 
    extrema, this can be done much more elegantly, e.g.: 

    if LoGI(2, j, i) == min(min(LoGI(:, j-1:j+1, i-1:i+1)))
    or for absolute local extrema, it could be done like this:
    if all(all(LoGI(2, j, i) < LoGI(:, j-1:j+1, i-1:i+1)))


(b) The most common mistake here was the lack of discussion about the limitations
    of the technique itself. What can go wrong when trying to estimate shape from the
    scale-space distribution of the features? Is the distribution planar in the 
    case of the shirt? 

(c) This question was generally well-done.