Efficient Structured Parsing of Facades Using Dynamic Programming

被引：52

作者：

Cohen, Andrea ^{[1
]}

Schwing, Alexander G. ^{[1
]}

Pollefeys, Marc ^{[1
,2
]}

机构：

[1] Swiss Fed Inst Technol, Zurich, Switzerland

[2] Univ Toronto, Toronto, ON M5S 1A1, Canada

来源：

2014 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR) | 2014年

关键词：

D O I：

10.1109/CVPR.2014.410

中图分类号：

TP18 [人工智能理论];

学科分类号：

081104 ; 0812 ; 0835 ; 1405 ;

摘要：

We propose a sequential optimization technique for segmenting a rectified image of a fac, ade into semantic categories. Our method retrieves a parsing which respects common architectural constraints and also returns a certificate for global optimality. Contrasting the suggested method, the considered fac, ade labeling problem is typically tackled as a classification task or as grammar parsing. Both approaches are not capable of fully exploiting the regularity of the problem. Therefore, our technique very significantly improves the accuracy compared to the state-of-the-art while being an order of magnitude faster. In addition, in 85% of the test images we obtain a certificate for optimality.

引用

页码：3206 / 3213

页数：8