Nonparametric scene parsing in the images of buildings. (August 2018)

• Record Type: Journal Article
• Authors: Talebi, Mehdi; Vafaei, Abbas; Monadjemi, S. Amirhassan
• Is Part Of: Computers & electrical engineering. Volume 70(2018)
• Page Start: 777
• Page End: 788
• Publication Date: 2018-08
• Subjects: Nonparametric image parsing -- Semantic segmentation -- Object recognition -- Building and door detection; Computer engineering -- Periodicals; Electrical engineering -- Periodicals; Electrical engineering -- Data processing -- Periodicals; Ordinateurs -- Conception et construction -- Périodiques; Électrotechnique -- Périodiques; Électrotechnique -- Informatique -- Périodiques; Computer engineering; Electrical engineering; Electrical engineering -- Data processing; Periodicals; Electronic journals; 621.302854
• DOI: 10.1016/j.compeleceng.2018.01.004
• Languages: English
• ISSNs: 0045-7906

Abstract

Highlights: In this paper, we present a nonparametric scene parsing in the images of buildings. An image is parsed into regions of building, door, ground, sky, and other objects. We improve the door detection accuracy using low-level features and contextual cues. We have collected 3093 images of 40 different types of buildings. We achieve the per-pixel rate 82.4% and the average per-class accuracy of 73.8%. Abstract: In this paper, we present a nonparametric approach to parse an image into regions of building, door, ground, sky, and other possible objects (such as cars, people, and trees). In a nonparametric method, first, similar images to that of the test are retrieved from a labeled training dataset. Then, the labels are transferred from the superpixels of the retrieved images to their corresponding superpixels of the test image. Finally, the conceptual Markov random field model is utilized to increase the superpixel labeling accuracy. In addition, we propose a method to improve door detection accuracy using the line, color, texture, and contextual cues. We have collected 3093 images of 40 different types of buildings from the LabelMe and Sun datasets, consisting of skyscrapers, shops, houses, apartments, churches, and so on. Experimental results on the dataset show the effectiveness of our approach with promising results. Graphical abstract: