Spatial distribution feature for 3D indoor scene labelling

In this paper, we propose an innovative approach for indoor scene labelling through a Bayesian framework with a RGB-D camera. In our approach, with the depth information, we develop a novel spatial feature vector that derives from the combination of three oriental distributions by using eigenvector decomposition and sub-space combination to capture the structural feature of 3D scene. We use Gaussian Mixture Distribution to compute the conditional likelihood density of these features. Meanwhile, a 6-connect pair-wise model that accommodates the relationship of 3D location is designed and used for labelling through Markov Random Field. Our approach is evaluated on several challenging dataset and is shown to have great superiority and effectiveness.