HomeNet: Layout Generation of Indoor Scenes from Panoramic Images Using Pyramid Pooling
DescriptionWe propose HomeNet, an end-to-end approach to generate 3D layouts for indoor scenes. It employs a Fully Convolutional Network (FCN) along with pyramid pooling to predict the main structure of the room using only a single 360-degree panorama. When provided with the input image, the FCN outputs the boundary and corner maps, which are further optimized using a significantly faster algorithm than previous approaches. Using this information, the main structure of the room is obtained, which undergoes affine transformations to generate the 3D layout. We find that using global prior representation obtained through pyramid pooling helps improve accuracy. When evaluated on the PanoContext and 2D-3D Stanford dataset, we find our model is more accurate than state-of-the-art methods while also being faster.