SIFT function implementation for feature detection from an image. Our example classifies images for the following classes: airplanes, cars, dog, faces and keyboard.

We use the Bag of Word model where we generate codewords for each class, i.e. 20,000 features from the dog class. Once we generated the codewords for each classes, we can create a dictionary of 500 codeword and normalise it. Once we have 500 codewords we can use it to create a histogram for all the test and training images. The histogram bin represents a codeword/feature that will be used to help us classify or label the image based off the classes we mentioned.

How to run the different steps of the Bag of Word model (and the time it takes). Paste the python commands into console/terminal.

Tested on machine with following specs:

1. 10th Generation Intel Core i7-1065G7, Windows 10 Home 64bit, 16 GB RAM, Intel Iris Plus Graphics

• Extract SIFT descriptors from training and test images
• Stores as binary file ...descriptors.npy
• Takes X hours

python SIFT.py

• Generate the codebook descriptors by euclidean and sad
• Also generate the smaller codebook with cluster of 20
• Stores as binary file ...codebook.npy
• Takes x hours

python gen_codebook.py

• Loads binary codebook for test and training
• Generate histograms based on those normal and small codebook
• Stores as binary file ...histogram_euclidean.npy or ...histogram_euclidean_small.npy
• Takes < 15 minutes

python gen_histograms.py

• Classify all the test images and returns image and label
• Output all classes with 10 images and the classified label
• Input the following parameters for different result

optional arguments:
  -h, --help  show this help message and exit
  -e          use codebook generated using euclidean distance
  -s          use small codebook
  --training  classify training images

python classification_by_euclidean.py

• Classify all the test images and returns image and label
• Output all classes with 10 images and the classified label
• Input the following parameters for different result

optional arguments:
  -h, --help  show this help message and exit
  -e          use codebook generated using euclidean distance
  -s          use small codebook

python classification_by_intersection.py

• Visualise image keypoints that were assigned to the same code word in the disctionary of visual words
• Input the following parameters for different result

optional arguments:
  -h, --help  show this help message and exit
  -e          use codebook generated using euclidean distance
  -s          use small codebook

python visualise_same_word_patches.py