Intro to Computer Vision Problem Set 3

ECS 189G: Intro to Computer Vision
Problem Set 3
Instructions
1. Answer sheets must be submitted on SmartSite. Hard copies will not be accepted.
2. Please submit your answer sheet containing the written answers in a file named:
FirstName_LastName_PS3.pdf. Please put your name on the answer sheet.
3. Please submit your code and input/output images in a zip file named:
FirstName_LastName_PS3.zip. Your code and input images should all be saved in the main
directory. Your output images can be saved either in the main directory or in a sub-directory
called “output” in your main directory.
4. You may collaborate with other students. However, you need to write and implement your
own solutions. Please list the names of students you discussed the assignment with.
5. For the implementation questions, make sure your code is documented, is bug-free, and
works out of the box. Please be sure to submit all main and helper functions. Be sure to not
include absolute paths. Points will be deducted if your code does not run out of the box.
6. If plots are required, you must include them in your answer sheet (pdf) and your c ode must
display them when run. Points will be deducted for not following this protocol.
1 Short answer problems [15 points]
1. When performing interest point detection with the Laplacian of Gaussian, how would
results differ if we were to (a) take any positions that are local maxima in scale-space, or (b)
take any positions whose filter response exceeds a threshold? Specifically, what is the impact
on repeatability or distinctiveness of the resulting interest points?
2. What exactly does the value recorded in a single dimension of a SIFT keypoint descriptor
signify?
3. If using SIFT with the Generalized Hough Transform to perform recognition of an object
instance, what is the dimensionality of the Hough parameter space? Explain your answer.
2 Programming: Video search with bag of visual words [85 points]
For this problem, you will implement a video search method to retrieve relevant frames from a video
based on the features in a query region selected from some frame. We are providing the image data and
some starter code for this assignment.
Provided data
You can access pre-computed SIFT features here:
• /usr/local/189data/sift/ (accessible from CSIF machines)
or
• \\coe-itss-bfs.engr.ucdavis.edu\Classdata\ECS189\Materials\sift\ (accessible from COE machines
in your T drive; e.g., Academic Surge lab machines)
The associated images are stored here:
• /usr/local/189data/frames/ (accessible from CSIF machines)
or
• \\coe-itss-bfs.engr.ucdavis.edu\Classdata\ECS189\Materials\frames\ (accessible from COE
machines in your T drive; e.g., Academic Surge lab machines)
Please note the data takes about 3 GB, so you should *not* try to copy it to your home directory. Just
point to the data files directly in your code. If you would like to download the data (e.g., to work on your
own machine), they are also available here:
SIFT features: https://ucdavis.box.com/s/4rrs7bp6cszrg6xrb6cr6v2mupnemwgu
Associated Images: https://ucdavis.box.com/s/0pn2bmlfr98o8hs3yjhjdgiluz9lvald
Each .mat file in the provided SIFT data corresponds to a single image, and contains the following
variables, where n is the number of detected SIFT features in that image:
descriptors nx128 single // SIFT vectors as rows
imname 1x57 char // name of image file that goes with this data
numfeats 1x1 single // number of detected features
orients nx1 single // orientations of the patches
positions nx2 single // positions of the patch centers
scales nx1 single // scales of the patches
Provided code
The following are the provided code files. You are not required to use any of these functions, but you will
probably find them helpful. You can access the code here:
http://web.cs.ucdavis.edu/~yjlee/teaching/ecs189g-spring2015/psets/pset3/pset3code.zip
• loadDataExample.m: Run this first and make sure you understand the data format. It is a
script that shows a loop of data files, and how to access each descriptor. It also shows how to
use some of the other functions below.
• displaySIFTPatches.m: given SIFT descriptor info, it draws the patches on top of an
image
• getPatchFromSIFTParameters.m: given SIFT descriptor info, it extracts the image
patch itself and returns as a single image
• selectRegion.m: given an image and list of feature positions, it allows a user to draw a
polygon showing a region of interest, and then returns the indices within the list of
positions that fell within the polygon.
• dist2.m: a fast implementation of computing pairwise distances between two matrices for
which each row is a data point
• kmeansML.m: a faster k-means implementation that takes the data points as columns
What to implement and discuss in the write-up
Write one script for each of the following (along with any helper functions you find useful), and in your
pdf writeup report on the results, explain, and show images where appropriate. Your code must access the
frames and the SIFT features from subfolders called ‘frames’ and ‘sift’, respectively, in your main working
directory.
1. Raw descriptor matching [20 pts]: Allow a user to select a region of interest (see
provided selectRegion.m) in one frame, and then match descriptors in that region to
descriptors in the second image based on Euclidean distance in SIFT space. Display the selected
region of interest in the first image (a polygon), and the matched features in the second image,
something like the below example. Use the two images and associated features in the provided
file twoFrameData.mat (in the zip file) to demonstrate. Note, no visual vocabulary should
be used for this one. Name your script rawDescriptorMatches.m
2. Visualizing the vocabulary [20 pts]: Build a visual vocabulary. Display example image patches
associated with two of the visual words. Choose two words that are distinct to illustrate what
the different words are capturing, and display enough patch examples so the word content is
evident (e.g., say 25 patches per word displayed). See provided helper function
getPatchFromSIFTParameters.m. Explain what you see. Name your script
visualizeVocabulary.m. Please submit your visual words in a file called kMeans.mat.
This file should contain a matrix of size kx128 called kMeans.
3. Full frame queries [20 pts]: After testing your code for bag-of-words visual search, choose 3
different frames from the entire video dataset to serve as queries. Display each query frame
and its M=5 most similar frames (in rank order) based on the normalized scalar product
between their bag of words histograms. Explain the results. Name your script
fullFrameQueries.m
4. Region queries [25 pts]: Select your favorite query regions from within 4 frames (which may be
different than those used above) to demonstrate the retrieved frames when only a portion of the
SIFT descriptors are used to form a bag of words. Try to include example(s) where the
same object is found in the most similar M frames but amidst different objects or
backgrounds, and also include a failure case. Display each query region (marked in the
frame as a polygon) and its M=5 most similar frames. Explain the results, including possible
reasons for the failure cases. Name your script regionQueries.m
Tips: overview of framework requirements
The basic framework will require these components:
• Compute nearest raw SIFT descriptors. Use the Euclidean distance between SIFT
descriptors to determine which are nearest among two images’ descriptors. That is, “match”
features from one image to the other, without quantizing to visual words.
• Form a visual vocabulary. Cluster a large, representative random sample of SIFT
descriptors from some portion of the frames using k-means. Let the k centers be the visual
words. The value of k is a free parameter; for this data something like k=1500 should work,
but feel free to play with this parameter [see Matlab’s kmeans function, or provided
kmeansML.m code]. Note: you may run out of memory if you use all the provided SIFT
descriptors to build the vocabulary.
• Map a raw SIFT descriptor to its visual word. The raw descriptor is assigned to the nearest
visual word. [see provided dist2.m code for fast distance computations]
• Map an image’s features into its bag-of-words histogram. The histogram for image I j is a kdimensional vector: F (I j ) = [ freq1,j , freq2,j , … , freqk,j], where each entry freqi,j counts the
number of occurrences of the i-th visual word in that image, and k is the number of total words
in the vocabulary. In other words, a single image’s list of n SIFT descriptors yields a kdimensional bag of words histogram. [Matlab’s histc is a useful function]
• Compute similarity scores. Compare two bag-of-words histograms using the normalized
scalar product.
• Sort the similarity scores between a query histogram and the histograms associated with the
rest of the images in the video. Pull up the images associated with the M most similar examples.
[see Matlab’s sort function]
• Form a query from a region within a frame. Select a polygonal region interactively with
the mouse, and compute a bag of words histogram from only the SIFT descriptors that fall
within that region. Optionally, weight it with tf-idf. [see provided selectRegion.m code]
3 OPTIONAL: Extra credit (up to 10 points each, max 20 points total)
• Stop list and tf-idf. Implement a stop list to ignore very common words, and apply tf-idf
weighting to the bags of words. Discuss and create an experiment to illustrate the impact on your
results.
• Spatial verification. Implement a spatial consistency check to post-process and re-rank the
shortlist produced based on the normalized scalar product scores. Demonstrate a query
example where this improves the results.