A Unified Model for Using the Higher-order Information in Semantic Segmentation Tasks

In this paper we propose a unified model for exploiting independent tasks of: Object recognition and Scene classification in Semantic segmentation task. These independent tasks is used as higher-order information in the Conditional Markov Random Field (CRF) framework. Our main contribution is constructing an structure for the CRF in combining aforementioned independent modules and defining resulting energy function for the CRF. Another contribution of this paper is implementing a heuristic approach for scene classification module in our problem. Recent researches in deep learning methods have shown promising results in many area of computer vision. In this paper we have used features extracted from Convolutional Neural Network in the object recognition and scene classification as two independent module. We have shown improvement results by adding these higher-order information to the model in semantic segmentation task on the two challenging datasets of 21-MSRC and Stanford Background dataset.

Keywords:

Semantic Segmentation , Conditional Random Fields , Convolutional Neural Networks , Object Detection , Scene Classification