/imagenet_metaclasses

Contains code to evaluate the WordNet synset tree and merge classes of ImageNet based on this information

Primary LanguagePythonMIT LicenseMIT

Create Imagenet Metaclasses

Use parse.py to create metaclasses to ImageNet by automatically merging classes along the paths defined by WordNet. Requires the files data.noun from WordNet and either imagenet_synsets.txt (ls ILSVRC2012_train > imagenet_synsets.txt) or imagenet_counts (use count_imagenet.py and point to the ImageNet Folder), all three are provided here.

Outputs the following files:

  • out.yaml the merged version of the synset tree.
  • class_dict.yaml a dictionary mapping the old classes to the new merged classes.
  • leaf_info.txt the resulting classes ordered by size.

Here is an example of the resulting sizes for a "bottom-up" merge