Hierarchical GoogLeNet
- Hierarchical Approach in Large Scale Image Recognition with Deep Learning
- ILSVRC 2015 - Miletos Team Entry (Results : http://image-net.org/challenges/LSVRC/2015/results)
- Took the 19th place in ILSVRC 2015 using without model and scale ensembling
- Took the 1st place in several classes (red fox, brown bear, ...) in classification track
Abstract
-
Hierarchical GoogLeNet is an adaptation of the GoogLeNet to use hierarchical category structure in the data for ILSVRC – Localization competition. Architecture of the Hierarchical GoogLeNet is the same as architecture of plain GoogLeNet.
-
There is a hierarchical structure in object categories and this hierarchical structure of the data is compatible with WordNet Tree. In WordNet Tree, objects are hierarchically categorised into a tree structure.
-
Hierarchical GoogLeNet is trained by considering this hierarchical structure of categories.