Nonparametric RGB-D Scene Parsing

With the development of depth or range sensors, the depth map has been publicly available and widely used in the field of computer vision in recent years, and RGB-D scene parsing also has received the widespread attention. However, the existing RGB-D scene parsing algorithms are mostly parametric models based on training method. In this paper, an effective nonparametric method based upon the label trans-ferring scheme is proposed for RGB-D scene parsing.This thesis first proposes a label transfer algorithm in markov random field framework, which includes label pool construction, bi-directional superpixel match-ing, and label transferring stages. Then a collaborative representation based classifi-cation (CRC) mechanism is built for Markov Random Field (MRF), and parsing result is achieved through minimizing the energy function via Graph Cuts. The effectiveness of our approach is validated on the indoor NYU Depth V1 dataset, showing that it outperforms both state-of-the-art RGB-D parsing techniques and a classical nonpa-rametric superparsing method.In addition, this thesis proposes acollective scene parsing algorithm based on semantic graph model. Similar images are obtained by image clustering, and then a k-nearest neighbor semantic graph, a L1 norm graph and a CRC graph are built for these similar images. Finally, a smooth labeling result is obtained by Graph Cuts. Expe-riments prove that the semantic graph achieves better accuracy and has better scala-bility.