Semantic Segmentation of Indoor-Scene RGB-D Images Based on Iterative Contraction and Merging

In this paper, we propose an iterative contraction and merging framework (ICM) for semantic segmentation in indoor scenes. Given an input image and a raw depth image, we first derive the dense prediction map from a convolutional neural network (CNN) and a normal vector map from the depth image. By combining the RGB-D image with these two maps, we can guide the ICM process to produce a more accurate hierarchical segmentation tree in a bottom-up manner. After that, based on the hierarchical segmentation tree, we design a decision process which uses the dense prediction map as a reference to make the final decision of semantic segmentation. Experimental results show that the proposed method can generate much more accurate object boundaries if compared to the state-of-the-art methods.