Image-Retrieval
Image Retrieval using methods like Vocabulary-Tree & Bag-Of-Words
Image retrieval refers to task of retrieving the most similiar image from dataset and is a very important problem in Computer-Vision. Due to massive size of dataset we have removed dataset from the repository but rest of code works.
Requirements
scipy
python-opencv
Bag Of Words (BOW)
The Code is implemented in the file BOW.py
- First features are extracted from the images
- Then for these features K Means clustering is performed with 1000 clusters.
- Features are aligned to closest cluster centers
- Histograms are created for every image mapping them to cluster centers corresponding to their images
- While querying we build a corresponding histogram for the test image and find the most similiar histogram from the training images (via a simple dot product)
Vocublary Tree
The Code is implemented in the file VocTree.py
- First the features are extracted from images
- The a tree is built recursively. The tree is represented via a
Nodeobject implemented in functions/Node.py. The buildTree function is implemented in functions/Utils.py - Then images and leafs are then matched to each other building hash-maps from both sides.
- For every image, we use leaves as codebooks/histogram and determine how much each leaf weighs in for an image. This is implememented in lines 51-64 of VocTree.py
- While querying we build a corresponding leaf based histogram for the test image and find the most similiar histogram from the training images (via a simple dot product)