<aside> ‼️ ImageNet and COCO datasets each having specific and general labels makes it hard to train YOLOv2 with both datasets. We need a coherent way to merge these labels.

</aside>

Word Tree

WordNet

• A language database that structures the concepts and how they relate
• A directed graph

WordTree

• A hierarchical tree from the concepts in ImageNet

• Making a wordtree
  • root node : ‘physical object’
  • add nodes(labels in ImageNet) that has only one path to the root node in wordnet
  • add rest of the nodes (labels that has more than one path to the root node in wordnet) in a way that would grow the wordtree as least as possible

Prediction with WordTree

Classification

• To preform classification we predict conditional probabilities at every node for the probability of each hyponym of that synset given that synset.

<aside> ❓ hyponym : 하위어 synset : set of synonym(동의어)

</aside>

example) At the “terrier” node we predict :

$P_r$(Norfolk terrier | terrier)

$P_r$(Yorkshire terrier | terrier)

$P_r$(Bedlington terrier | terrier)

               $\\ldots$

• To compute absolute probability for a particular node we follow through path through the tree to the root node and multiply to conditional probablities.