Indoor Scene 3D Layout and Clutter Estimation from RGB-D Images
Mehdi Noroozi, Mostafa Kamali Tabrizi, Seyed Reza Moghadasi · 2014
We propose a new method for recovering indoor scene layout and clutter from RGB-D images. A robust method is introduced for obtaining scene coordinates from depth images by solving a restricted quantization problem over normal vectors. Using this quantization and 3D position of pixels we segment the image to planar surfaces and obtain an orientation labeling. The segmented image is used to extract features for layout candidates generated by sampled rays from vanishing points. These features are applied in a structured learning algorithm to rank layout candidates. Our approach recover clutter by distinguishing different layers of parallel surfaces. Our experimental results on the challenging NYU v2 dataset show that our approach outperforms state-of-the-art methods.