Abstract
Abstract. In this paper we explore semantic segmentation of man-made scenes using fully connected conditional random field (CRF). Images of man-made scenes display strong contextual dependencies in the spatial structures. Fully connected CRFs can model long-range connections within the image of man-made scenes and make use of contextual information of scene structures. The pairwise edge potentials of fully connected CRF models are defined by a linear combination of Gaussian kernels. Using filter-based mean field algorithm, the inference is very efficient. Our experimental results demonstrate that fully connected CRF performs better than previous state-of-the-art approaches on both eTRIMS dataset and LabelMeFacade dataset.
Highlights
Semantic segmentation of man-made scenes is one of the fundamental problems in photogrammetry and computer vision
In contrast with local-range conditional random field (CRF) (Rother et al, 2004, Shotton et al, 2006), which are solved by an expensive discrete optimization problem (Kappes et al, 2015), mean field approximation inference for the fully-connected CRF is much more efficient (Krahenbuhl and Koltun, 2011)
We propose to use fully connected CRF to model semantic segmentation of man-made scene problem and demonstrate it leads to state-of-the-art results
Summary
Semantic segmentation of man-made scenes is one of the fundamental problems in photogrammetry and computer vision. E.g. street scene, as shown, may be the most familiar scenes in our life. Applications of man-made scene interpretation include 3D city modeling, vision-based outdoor navigation, intelligent parking etc. Man-made scenes exhibit strong contextual and structural information in the form of spatial interactions among components, which may include buildings, cars, doors, pavements, roads, windows or vegetation. The eTRIMS (Korcand Forstner, 2009) and LabelMeFacade (Frohlich et al, 2010, Brust et al, 2015) image databases are two popular dataset for man-made scene semantic segmentation, which have irregular facades and do not follow strong architectural principles. We will explore semantic segmentation of this kind of man-made scenes using fully connected Conditional random fields (CRFs)
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
