1. ImageNet 10K Dataset

ImageNet dataset is constructed based on the WordNet hierarchy and it is under development. The latest release of ImageNet contains 14 million images from 21841 classes.



ImageNet10K contains more than 9 million images from 10179 classes. These classes might be parent or leaf nodes in the hierarchy. In the leaf nodes, the images show instances of a specific entity. The Top-1 classification accuracy of leaf nodes are generally higher than the parent nodes due to this cleanness of the images.

5 Classes10 Classes20 Classes with Top Accuracy75% Accuracy50% Accuracy25% Accuracy10% Accuracy5% Accuracy0% Accuracy

2. ImageNet Large Scale Visual Categorization Challenge Dataset 2010 (ILSVRC2010)

The 2010 challenge dataset contains 1000 classes and 1.2M images. The whole hierarchy contains 1676 nodes, 1000 of which are leaf nodes. A child node may have at most 3 parents therefore the hierarchy is an undirected graph but without any loops.

To VISUALIZE THE HIERARCHY in ILSVRC2010 dataset, we use OpenOrd method ^[*] which is based on Fruchterman-Reingold algorithm and works with fixed number of iterations controlled via a simulated annealing type schedule.

• Click a node in the graph, its immediate neighbours will be
  • listed on the right side
  • highlighted on the graph.
• Remove/replace the nodes and edges that are not connected to a particular node.
• Zoom in/out of the whole graph.
• Search a particular node in the search box at the top of the page.
• Use magnifying glass to zoom a particular node on the graph.
• The size of the nodes are determined with the number of children they have.
• The root node can be searched by the name 'entity'.