[SOLVED] (csc420) assignment 2 1. interest point detection: (a) [2 points] write your own function for harris corner metric

1. Interest point detection:
(a) [2 points] Write your own function for Harris corner metric using the harmonic
mean (slide 29, lecture 6). Display your result for the attached image building.jpg showing your cornerness metric output. You can use built-in functions
for convolution, gradients, but you must compute M yourself. Adjust α to get a
good result.(b) [2 points] Write your own function to perform non-maximal suppression using
ordfilt2.m or your own morphological operators function of choice. Use a circular
element, and experiment with varying radii r as a parameter. Explain why/how
the results change with r.(c) [2 points] Write code to search the image for scale-invariant interest point (i.e.
blob) detection using the Laplacian of Gaussian and checking a pixel’s local neighbourhood as in SIFT. You may use code from tutorial 4 as a starting point. You
must find extrema in both location and scale. Find the appropriate parameter
settings, and display your keypoints for synthetic.png. Hint: Only investigate
pixels with the LoG above or below a threshold.(d) [1 point] Compare and contrast the Harris corner metric with non-maximal supression as a keypoint detector to the Laplacian of Gaussian method. Show examples where they detect different keypoints and the same keypoints and explain
why they are the same/different using synthetic.png and building.png.2. For this question you will use interest point detection for matching using SIFT. You
may use a SIFT implementation (e.g. http://www.vlfeat.org/), or another, but specify
what you use.(a) [0.5 points] Extract SIFT keypoints and features for book.jpg and findBook.jpg.(b) [1.5 points] Write your own matching algorithm to establish feature correspondence between the two images using the reliability ratio on Lecture 8, slide 23.
You can use pdist2.m, but you must find the matches yourself. Experiment for
different thresholds.(c) [2 points] Affine transformation: Use the top k correspondences from part (b)
to solve for the affine transformation between the features in the two images via
least squares using the Moore-Penrose psudeo inverse. Demonstrate your results
for various k. Use only basic linear algebra libraries.(d) [0.5 point] Visualize the affine transformation. Do this visualization by taking
the four corners of the reference image, transforming them via the computed affine
transformation to the points in the second image, and plotting those transformed
points. Please also plot the edges between the points to indicate the parallelogram.
If you are unsure what the instruction is, please look at Figure 12 of [Lowe, 2004].(e) [1.5 points] Write code to perform matching that takes the colour in the images
into account during SIFT feature calculation and matching. Explain the rational behind your approach. Use colourTemplate.png and colourSearch.png,
display your matches with (2.d).3. Extra: [3 points] total (this is an optional exercise) Implement your own SURF feature descriptor, which is very similar in steps to SIFT. You can use any keypoint localization algorithm of your choice. See: http://www.vision.ee.ethz.ch/ surf/eccv06.pdf
and or Algorithm 6 in http://www.ipol.im/pub/art/2015/69/.You may submit up to two days late without penalty if you do well in this
exercise.
(a) [2 points] Your descriptor must include:
• Build your own Harr-wavlet filters
• Computation of Harr-wavlet responses
• Obtain the orientation using a circular neighbourhood of radius 6s.• Gaussian weighting
• Rotation and scale invariance via feature window re-orientation and scale
fitting (you can use bilinear interpolation instead of integral images)
• Feature sub-binning into a 4×4 block of regions and 16 features per region(b) [1 points] Compare your algorithm to SIFT for the book.jpg and findBook.jpg
pair, and at least two additional of your own image pairs, and discuss the pros/cons
in accuracy. Don’t worry about the speed of your implementation.