Description
1. (30 points) Harris Corner Detection.
Input image: 2/data/boat.mat.
Assume the pixel dimensions to be equal along both axes, i.e., assume an aspect ratio of 1:1 for
the axes.
Shift and rescale the intensities in the image to lie within the range [0, 1].
Implement the Harris corner detector algorithm. The parameters underlying this algorithm are:
two Gaussian smoothing levels involved in computing the structure tensor, the constant k in the
corner-ness measure. Tune these three parameters to get the best results.
• Write a function myHarrisCornerDetector.m to implement this.
• Display the derivative images, corresponding to the derivatives along the X and Y axes.
• Display the images (along with a colormap) of the two eigenvalues of the structure tensor,
evaluated at each pixel.
• Display the image (along with a colormap) of the Harris corner-ness measure. Positive values
in this image must correspond to a corner structure in the image.
• Report all three parameter values used.
2. (40 points) Image Segmentation using mean shift.
Input image: 3/data/baboonColor.png.
Take this 512 × 512 pixel image, smooth it using Gaussian convolution with standard deviation
1 pixel width, and subsample the smoothed image by a factor of 2 in each spatial dimension to
produce a 256 × 256 image. Use this smaller-sized image for the following experiment. If this
image is still too large for your computer’s memory, then you may resize further.
1
Implement the algorithm for mean-shift image segmentation using both color (RGB) and spatialcoordinate (XY) features. Tune parameters suitably to get a segmented image with at least 5
segments and no more than 50 segments. To improve code efficiency, you may use Matlab
functions like knnsearch(), bsxfun(), etc. For this image, about 20 iterations should be sufficient
for reaching close to convergence. You may select a random subset of nearest neighbors, in
feature space, for the mean-shift updates to reduce running time. Each iteration can run in about
10-20 seconds on a typical personal computer.
• Write a function myMeanShiftSegmentation.m to implement this.
• Display the (i) original image along with (ii) the segmented image that shows color-coded pixels
(and, thus, segments) using the color component of the converged feature vectors.
• Report the following parameter values: Gaussian kernel bandwidth for the color feature, Gaussian kernel bandwidth for the spatial feature, number of iterations.
2
