Daily CVPR - Wednesday

CVPR Daily: What is your work about? Kevin Chen: The work is about taking a single monocular RGB image, and estimating the layout from just that one image. So the layout means you have for every pixel label of which wall in the image the pixel belongs. This is only for indoor scenes. CVPR Daily: What is the novelty in this work? Kevin Chen: The novelty consists in a completely different approach from previous methods, which try to generate these layout candidates then use a structured SVM to try to rank the layout candidates proposals. Instead of doing that, we treat this as a semantic segmentation problem. We pass the input image to a fully convolutional neural network (FCNN) to generate an initial estimate of the layout. Then this layout doesn’t conform to a standard room because a standard room has straight lines on the edges. So to alleviate this, we combine the FCNN with a novel optimizer that predicts the layout which conforms to a normal room layout. 6 Presentations CVPR Daily: Wednesday Kuan Fang and Kevin Chen