Semantic Image Segmentation Using Random Forest and Single Class Histogram Model
Mentor: Dr. Amitabha MukherjeeRohan JingarMridul Verma
Problem Statement
The Problem statement is to segment the images using random forest using Single Histogram Class Model(SHCM) based on Textons as a Node-Test.We are studying a model for Semantic Segmentation of the image in different classes of objects. Segmentation means to label different connected regions of the image as belonging to specific object class. The system will be trained on set of images of some finite categories of objects. Some modification of Random Forest Classifier will be used for selecting features of the objects. The following figure explains what we want to achieve given a test image.
Our Contribution
We are following the work of F. Schroff et.al.[1] in which they have proposed a working model for Image Segmentation using Random Forest as a classifier which does the classification at pixel level. The code released by F. Schroff is able to train the classifier based on basic node test in RGB, HOG, F-17 feature space. The code for computing textons, texton-map, and building SHCM for each class was not present in the released version.So we build the Single Class Histogram Model based upon the Texton features and we were successfully able to add in the Released version.
Results
Code
| Code Released By F. Schroff: | RF_Seement v1.0 | |
| Dataset Used: | MSRC dataset | |
| Code Modified(SCHM + Texton as a feature pool): | Modified Code v1.0 , README | |
| Embedded Code(SCHM + Random Forest): | Embedded Code v1.0 , README |