基于马尔科夫随机场的非参数化RGB-D场景理解

Nonparametric RGB-D scene parsing based on Markov random field model

针对RGB-D场景下的场景理解问题,提出高效的基于标签传递机制的非参数化场景理解算法.该算法主要分为标签源构建、超像素双向匹配和标签传递三个步骤.与传统的参数化RGB-D场景理解方法相比,该算法不需要繁琐的训练,具有简单高效的特点.与传统的非参数化场景理解方法不同,该算法在系统的各个设计环节都有效利用了深度图提供的三维信息,在超像素匹配环节提出双向匹配机制,以减少特征误匹配;构建基于协同表示分类(CRC)的马尔科夫随机场(MRF),用Graph Cuts方法求出最优解,获得场景图像每个像素的语义标签.该算法分别在室内的NYU-V1数据集和室外的KITTI数据集上进行实验.实验结果表明,与现有算法相比,该算法取得了显著的性能提升,对室内、外场景均适用.

An effective nonparametric method was proposed for RGB-D scene parsing. The method is basedupon the label transferring scheme, which includes label pool construction, bi-directional superpixel matchingand label transferring stages. Compared to traditional parametric RGB-D scene parsing methods, theapproach requires no tedious training stage, which makes it simple and efficient. In contrast to previousnonparametric techniques, our method not only incorporate geometric contexts at all the stages, but alsopropose a bi-directional scheme for superpixel matching in order to reduce mismatching. Then a collaborativerepresentation based classification （CRC） mechanism was built for Markov random field （MRF） , andparsing result was achieved through minimizing the energy function via Graph Cuts. The effectiveness ofthe approach was validated both on the indoor N Y U Depth V I dataset and the outdoor K ITTI dataset. Theapproach outperformed both state-of-the-art RGB-D parsing techniques and a classical nonparametric superparsingmethod. The algorithm can be applied to different scenarios, having a strong practical value.