Scribd
Upload
0 views

DM-Ass03

Uploaded by

i200847 Fatima Asim
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
0 views

DM-Ass03

Uploaded by

i200847 Fatima Asim
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 5

Assignment 3 DS3002-Data Mining Total marks: 100

DS3002-Data Mining
Assignment #3
Spring 2024

Total: 100
Deadline: 30 March 2024

1|Pag
e
Assignment 3 DS3002-Data Mining Total marks: 100

Important Instructions
 You must complete the assignment before the deadline.
Since the assignment is a little lengthy, get started right
away!

 Plagiarism is strictly prohibited! Try to solve the


assignment on your own. Marks will be given based on
your thought process so make sure you have a solid
reason for all your attempts.

 No assignment will be accepted after the due date.

 If Plagiarism Found, Straight ZERO in Whole Assignment.

Deliverables:

 Submission Format: i21-XXXX_A3.zip


 Submit 2 Jupyter Notebooks (Q1 and Q2 Separately) and
Report
Marks Distribution

 Report 10 Marks
 Q1 (50 Marks)
 Q2 (5+20+5+10 = 40 Marks)

2|Pag
e
Assignment 3 DS3002-Data Mining Total marks: 100

Q1. Interactive Foreground Segmentation Using K-Means Clustering


You will implement a basic version of the interactive image cut-out / segmentation
approach called Lazy Snapping [1]. You are given several test images, along with
corresponding auxiliary images depicting the foreground and background “seed” pixels
marked with red and blue brush-strokes respectively. Your program should exploit these
partial human annotations in order to compute a precise figure-ground segmentation.

You should first write a function that performs K-Means Clustering on color pixels. The
input arguments are the desired number of clusters k and the data points to cluster. It
outputs a cluster index for each input data point, as well as the k cluster centroids. You
should then extract the seed pixels for each class (i.e., foreground and background) and
use your k-means function to obtain N clusters for each class. A good choice for N is 64,
but you can experiment with smaller or bigger values

Next, compute the likelihood of a given pixel p to belong to each


of the N clusters of a given class (either foreground or
background) using an exponential function of the negative of the
Euclidean distance between the respective cluster center 𝐶𝑘 and
the given pixel 𝐼𝑝 in the RGB color space. The overall likelihood
𝑝(𝑝) of the pixel to belong to this class is a weighted sum of all
these cluster. For this Assignment 𝑤𝑘 values will be 0.1

You might see improved results if you remove the square operation in equation above.
The reason may have to do with the fact that squaring essentially reduces small values
even further, mitigating (adversely) the effects of the negative exponential. Finally, a
given pixel is simply assigned to the class to which it is more likely to belong. That is, if
𝑝𝑓𝑔(𝑝) > 𝑝𝑏𝑔(𝑝), pixel 𝑝 is assigned to the foreground class, i.e., 𝑥𝑝= 1, and vice versa.

Include your results for all test images in your report, and explain what you get. For test
images with two stroke images, you should report results for both cases. Also compare
results for different values of N, i.e., the number of clusters evolved in the foreground
and background classes.

Results for N=64 (both foreground and background)

3|Pag
e
Assignment 3 DS3002-Data Mining Total marks: 100

Q2. Face Recognition Using K-NN


For this question, we use simplified version of CMU Pose, Illumination, and Expression
(PIE) Dataset, which only contains 10 subjects spanning five near-frontal poses, and there
are 170 images for each individual. In addition, all the images have been resized to
32x32 pixels. The dataset is provided in the form of a csv file with 1700 rows and 1024
columns. Each row is an instance and each column a feature. The first 170 instances
belong to the first subject, the next 170 to the second subject and so on. Following
illustrate various instances of the first subject.

Tasks
1. Pre-process the dataset by normalizing each face image vector to unit
length (i.e., dividing each vector by its magnitude). Next, for each of the
10 subjects, randomly select 150 images for training and use the
remaining 20 for testing.

2. Implement a k Nearest Neighbors (k-NN) classifier from scratch and using


the training set and evaluate its performance on the test set. You may not
use built-in / library functions to implement the classifier.
a. You should also implement Euclidean and cosine similarity distance
measures and used them for different values of K.
b. You should also present results when fewer training images are
used (for instead 100 training images and 70 test images per
subject)
c. Write results in the report for k-values 2, 5, 7, and 11 with each distance
metric.

3. Use Sklearn to apply SVM and GaussianNB on this dataset and compare
the accuracy with K-NN in the report.

4. Perform dimensionality reduction on the training and testing datasets and


visualize them in 3-D space using the Principal Component Analysis (PCA)
and matplotlib or Seaborn.

4|Pag
e
Assignment 3 DS3002-Data Mining Total marks: 100

Rubric

Criteria Marks
Marks Distribution
- Report 10
- Q1 50
- Q2 40

Q1. Interactive Foreground Segmentation Using K-Means


- K-Means Clustering function implementation 15
- Seed pixel extraction and likelihood computation 20
- Experimentation and comparison with different N values 15

Q2. Face Recognition Using K-NN


- Pre-processing and data splitting 5
- k-NN classifier implementation 20
- Evaluation with different K values 10
- Evaluation with fewer training images 5
- Comparison with SVM and GaussianNB 5
- Dimensionality reduction and visualization 10

Overall Presentation
- Clarity and organization of the report 5
- Documentation and comments in code 5

5|Pag
e

You might also like