We used the Contextual Hierarchy Models (CHM) developed by Seyedhosseini ET. Al which learns contextual information in a hierarchical framework for semantic segmentation. At each level of the hierarchy, a classifier is trained based on down-sampled input images and output of previous levels. We then incorporate the resulting multidimensional contextual information into a classifier to segment the input image at original resolution. This training method allowed for optimization of a joint posterior probability at multiple resolutions through the hierarchy. To do this, we use CHM since they run on CPU’s in less time than Convolutional Neural Networks (CNN) due to its stage by stage training process. We also employ a segmentation framework that takes advantage of both input image and contextual information. This information is then learned in a Supervised framework making it more discriminative compared to auto-context algorithms as done by other groups. By using the required steps in employing a CHM, we later defined the bottom-up approach where multiple classifiers are learned and then use a top-down classifier to train one classifier at a time making it beneficial due to the prior knowledge developed through the bottom-up stage. This way we us simple filtering to create context images at different levels. We then used a Maximum Aposterirori (MAP) for knowing the posterior distribution directly. We eventually trained 8 CHM in a one-vs.-all architecture as proposed previously on the Stanford Background Dataset and produced improved pixel-wise and class-average accuracy as the baseline at that time.
Custom Dataset can be downloaded from the following link:- https://drive.google.com/drive/folders/0B7owL4cO7Q0_OUt1OXd3VG0wZ1k?usp=sharing
- All four folders need to be placed in the root directory.