Bag of Words Robotics

• Recognizing objects using local descriptors would be computationally expensive.
• The number of local features for a given object mainly depends on the size of the object, and therefore, varies for different objects.
• The key idea for fast 3D object recognition is to use mechanisms for representing objects in a compact and uniform format (e.g., histogram).
• If we represent objects in a uniform format, then we can apply ML algorithms
• Compute local features for all the discovered objects and make a pool of features.
• A dictionary is generated via Clustering of the pool of features into N clusters (the number of the clusters is the codebook size).
• Visual word are then defined as the centres of the extracted clusters.
• Finally, each object is described (abstracted) by a histogram of occurrences of these visual words.
• 
•