Skip to content

Latest commit

 

History

History
33 lines (21 loc) · 1.75 KB

File metadata and controls

33 lines (21 loc) · 1.75 KB

Reserach Design

  • Investigating whether Capsule Networks can provide useful formalism and structure for Class-Angostic Counting networks.
  • Fergus Steel
  • 2542391s
  • Dr. Paul Siebert

Hypothesis

  • Capsule Networks will increase the quality of the representations of visual concepts in Class Agnostic Counting

What is gained from this research?

  • Class-Agnostic counting theoretically allows for us to count objects in any object domain.
  • Capsule Networks are a relatively new area of reserach in Artificial Intelligence, and by demonstrating their use in a novel domain, we can make a valuable contribution to the literature by expanding the use cases of CapsNets.

How can we demonstrate this value?

  • Is a Class-Agnostic Capsule Network Architecture capable of predicting the density maps of visual concepts? – (1) representational general utility
  • Can Capsule Networks be used to count in unseen distributions/domains in Class Agnostic Counting Tasks? - (2) representational adaptability
  • Can the use of Encapsulated Representaions improve density-map estimation in Class Agnostic Counting? – (3) representational specific utility

How will we experiment this

  • Building a Capsule Network that system that attempts to show the above concepts,
    • (1) - To do this the network must be able to create density maps on trained distributions
    • (2) - To do this the network must be able to create density maps on object distributions that have not been seen
    • (3) - To do this the network must be able to demonstrate that it is an improvement on existing seminal research

What Technologies will be used?

  • FSC-147 dataset, with similiar method in Ranjan paper for passing in bounding boxes (i.e. a text file that is generated by GUI.)