Towards efficient support relation extraction from RGBD images

To extract reasonable support relations from “RGB + depth” (RGBD) images, it is very important to achieve good scene understanding. This paper proposes a novel approach to extracting accurate support relationships by analyzing the RGBD images of indoor scenes. Noting that the support relations and structure classes of indoor images are inherently related to physical stability, we construct an improved energy function that embodies this stability. We then infer the support relations and structure classes from indoor RGBD images by minimizing this energy function. Moreover, the authors succeed in improving the segmentation quality of RGBD images using the inferred results as input. Compared with previous methods, our approach produces more reasonable support relations and structure classes, where physical stability function is taken into account for resolving the optimization problem. We use the NYU-Depth2 dataset as the training data, and experimental results show that the proposed RGBD image segmentation method based on support relation abstraction produces more accurate results than segmentation methods based on ground-truth support relations.