Bag of Words

Common way to implement Loop Closure.

Resources

• Bag of Visual Words (Cyrill Stachniss)
• https://github.com/ovysotska/in_simple_english/blob/master/bag_of_visual_words.ipynb

I can trace it to this paper, which was cited by ORB-SLAM

• Bags of Binary Words for Fast Place Recognition in Image Sequences

In my own words:

1. Cluster together descriptors, since you have 2^256 possible descriptors, condense it down to groups of words through K-Means
   • this is the learning phase
2. Then, for each feature detected on the image, assign it to a cluster
3. Then, you can build a histogram of words, and use that as comparison between various images. If it is the same histogram, then you’re probably looking at the same image

This image helped me understand