Scene Parsing Based on A Two-Level Conditional Random Field

Recently,there have been increasing interests in semantic scene parsing,which are mainly based on prior models.However,as prior models only consider the common merits within objects,they ignore the inner color coherence lied in the single object and suffer from detail loss.In this paper,we utilize color cue as the inner information,and present an approach to incorporating local color model with prior models under a two-level conditional random field to preserve scene parsing details.More specially,objects of the scene are first roughly extracted using prior models within a superpixel-based conditional random field,in which prior models are acquired by supervised learning using color,gradient,texture and geometric cues,and then a local color model is built for each object based on the initial parsing result.Combining the local color model and prior models,we employ an EM(Expectation-Maximization)scheme for iterative refinement within a pixel-based conditional random field.Experimental evaluations with state-of-the-art methods verify that our approach is able to preserve details and achieve better performance.