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 labeled training images. Among two most popular techniques to address this are Random Forests (RF) and Neural Networks (NN). The main contribution work is explore relationship between special forms these techniques: stacked RFs deep Convolutional (CNN). show that there exists mapping from RF CNN, an approximate back. This insight gives major practical benefits: Firstly, CNNs can be intelligently constructed initialized, which crucial when dealing with limited amount data. Secondly, it utilized create new improved performance. Furthermore, yields CNN architecture, well suited for labeling. experimentally verify benefits different application scenarios in computer vision biology, where layout parts important: Kinect-based body part labeling depth images, somite microscopy images developing zebrafish.