Relating Cascaded Random Forests to Deep Convolutional Neural Networks for Semantic Segmentation

David L. Richmond, Dagmar Kainmueller, Michael Y. Yang, Eugene W. Myers, Carsten Rother

Published 2015-07-27Version 1

We consider the task of pixel-wise semantic segmentation given a small set of labelled training images. Among two of the most popular techniques to address this task are Random Forests (RF) and Neural Networks (NN). The main contribution of this work is to explore the relationship between two special forms of these techniques: cascaded RFs and deep Convolutional Neural Networks (CNN). We show that there exists an (approximate) mapping from cascaded RF to deep CNN, and back. This insight gives two major practical benefits: a) the performance of a greedily trained cascaded RF can be improved; b) a deep CNN can be intelligently constructed and initialized. Furthermore, the resulting CNN architecture has not yet been explored for pixel-wise semantic labelling. We experimentally verify these practical benefits for the task of densely labelling segments of the developing zebrafish body plan in microscopy images.