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

摘要

We consider the task of pixel-wise semantic segmentation given a small set oflabeled training images. Among two of the most popular techniques to addressthis task are Random Forests (RF) and Neural Networks (NN). The maincontribution of this work is to explore the relationship between two specialforms of these techniques: stacked RFs and deep Convolutional Neural Networks(CNN). We show that there exists a mapping from stacked RF to deep CNN, and anapproximate mapping back. This insight gives two major practical benefits:Firstly, deep CNNs can be intelligently constructed and initialized, which iscrucial when dealing with a limited amount of training data. Secondly, it canbe utilized to create a new stacked RF with improved performance. Furthermore,this mapping yields a new CNN architecture, that is well suited for pixel-wisesemantic labeling. We experimentally verify these practical benefits for twodifferent application scenarios in computer vision and biology, where thelayout of parts is important: Kinect-based body part labeling from depthimages, and somite segmentation in microscopy images of developing zebrafish.