CS 6643 Project 2: Human Detection using HOG Feature

Description

Write a program to compute the HOG (Histograms of Oriented Gradients)
feature from an input image and then classify the HOG feature vector into human or no-human by
using a 3-nearest neighbor (NN) classifier. In the 3-NN classifier, the distance between the input
image and a training image is computed by taking the histogram intersection of their HOG feature
vectors:
where I is the HOG feature of the input image and M is the HOG feature of the training image;
the subscript j indicates the jth component of the feature vector and n is the dimension of the
HOG feature vector. The distance between the input image and each of the training images is
computed and the classification of the input image is taken to be the majority classification of the
three nearest neighbors.
Conversion to grayscale: The inputs to your program are color images cut out from a larger
image. First, convert the color images into grayscale using the formula 𝐼𝐼 = 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅(0.299𝑅𝑅 +
0.587𝐺𝐺 + 0.114𝐵𝐵) where R, G and B are the pixel values from the red, green and blue channels
of the color image, respectively, and Round is the round off operator.
Gradient operator: Use the Prewitt’s operator for the computation of horizontal and vertical
gradients. Use the formula 𝑀𝑀(𝑖𝑖,𝑗𝑗) = �𝐺𝐺𝑥𝑥
2 + 𝐺𝐺𝑦𝑦
2 to compute gradient magnitude, where
𝐺𝐺𝑥𝑥 and 𝐺𝐺𝑦𝑦 are the horizontal and vertical gradients. Normalize and round off the gradient
magnitude to integers within the range [0, 255]. Next, compute the gradient angle. For image
locations where the templates go outside of the borders of the image, assign a value of 0 to both
the gradient magnitude and gradient angle. Also, if both 𝐺𝐺𝑥𝑥 and 𝐺𝐺𝑦𝑦 are 0, assign a value of 0 to
both gradient magnitude and gradient angle.
HOG feature: Refer to the lecture slides for the computation of the HOG feature. Use the
unsigned representation and quantize the gradient angle into one of the 9 bins as shown in the
table below. If the gradient angle is within the range [180, 360), simply subtract the angle by 180
first. Use the following parameter values in your implementation: cell size = 8 x 8 pixels, block
size = 16 x 16 pixels (or 2 x 2 cells), block overlap or step size = 8 pixels (or 1 cell.) Use L2
norm for block normalization. Leave the histogram and final feature values as floating point
numbers (Do not round off to integers.)
Histogram Bins
Bin # Angle in degrees Bin center
1 [0,20) 10
2 [20,40) 30
3 [40,60) 50
4 [60,80) 70
5 [80,100) 90
6 [100,120) 110
7 [120,140) 130
8 [140,160) 150
9 [160,180) 170
CS 6643 E. K. Wong
Project 2: Human Detection using HOG Feature Fall 2021
Training and test images: A set of 20 training images and a set of 10 test images in .bmp format
will be provided. The training set contains 10 positive (human) and 10 negative (no human)
samples and the test set contains 5 positive and 5 negative samples. All images are of size 160
(height) X 96 (width). With the given image size and the parameters given above for computing
the HOG feature, there are 20 X 12 cells and 19 X 11 blocks in the detection window. The
dimension of the HOG feature vector is 7,524.
Implementation: You can use Python, C++/C, Java or Matlab to implement your program. If
you would like to use a different language, send me an email first. You are not allowed to use any
built-in library functions for any of the tasks you are required to implement, including the
Prewitt’s operator and the computation of HOG features and histogram intersections. The only
library functions you are allowed to use are those for the reading and writing of image files,
matrix and vector arithmetic, and other commonly used mathematical functions.
Hand-in: Hand in the following files on BrightSpace by the due date. Please submit as separate
files, do not ZIP.
• Your source code file. Put comments in your source code to make it easier for someone
else to read your program. Points will be taken off if you do not have comments in your
source code.
• The ASCII (.txt) files containing the HOG feature values for three of the training images
and three of the test images (one file per image.) I will let you know which images to
provide HOG feature values later. The feature values should be separated by line breaks.
You should have 7,524 lines in each file.
• A PDF report that contains the following:
o Instruction on how to run your program and instruction on how to compile your
program if your program requires compilation.
o Normalized gradient magnitude images for the 10 test images.
o The source code of your program (Copy-and-paste from source code file. This is
in addition to the source code file that you need to hand in.)
o Use the table below to report the classification results. Distance is computed by
using the histogram intersection formula above.
Test image Correct
Classification
File name of 1st
NN, distance &
classification
File name of 2nd
NN, distance &
classification
File name of 3rd
NN, distance &
classification
Classification
from 3-NN
crop001034b Human
crop001070a Human
crop001278a Human
crop001500b Human
person_and_bike_151a Human
00000003a_cut No-human
00000090a_cut No-human
00000118a_cut No-human
no_person_no_bike_258_
cut
No-human
no_person_no_bike_264_
cut
No-human