Global Stereo Reconstruction under Second-Order Smoothness Priors

被引：148

作者：

Woodford, Oliver ^{[1
]}

Torr, Philip ^{[2
]}

Reid, Ian ^{[1
]}

Fitzgibbon, Andrew ^{[3
]}

机构：

[1] Univ Oxford, Dept Engn Sci, Parks Rd, Oxford OX1 3PJ, England

[2] Oxford Brookes Univ, Dept Comp, Oxford OX33 1HX, England

[3] Microsoft Res Ltd, Cambridge CB3 0FB, England

来源：

IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE | 2009年 / 31卷 / 12期

基金：

英国工程与自然科学研究理事会;

关键词：

Stereo; second-order prior; discrete optimization; graph cuts; MARKOV RANDOM-FIELDS; ENERGY MINIMIZATION; BAYESIAN-APPROACH;

D O I：

10.1109/TPAMI.2009.131

中图分类号：

TP18 [人工智能理论];

学科分类号：

081104 ; 0812 ; 0835 ; 1405 ;

摘要：

Second-order priors on the smoothness of 3D surfaces are a better model of typical scenes than first-order priors. However, stereo reconstruction using global inference algorithms, such as graph cuts, has not been able to incorporate second-order priors because the triple cliques needed to express them yield intractable (nonsubmodular) optimization problems. This paper shows that inference with triple cliques can be effectively performed. Our optimization strategy is a development of recent extensions to alpha-expansion, based on the "QPBO" algorithm. The strategy is to repeatedly merge proposal depth maps using a novel extension of QPBO. Proposal depth maps can come from any source, for example, frontoparallel planes as in alpha-expansion, or indeed any existing stereo algorithm, with arbitrary parameter settings.

引用

页码：2115 / 2128

页数：14