In this paper, we aim to understand the semantics and 3D structure of a scene from a single depth image. Recent deep neural networks based methods aim to simultaneously learn object class labels and infer the 3D shape of a scene represented by a large voxel grid. However, individual objects within the scene are usually only represented by a few voxels leading to a loss of geometric detail. In addition, significant computational and memory resources are required to process the large scale voxel grid of a whole scene. To address this, we propose an efficient and holistic pipeline, 3R-Depth, to simultaneously learn the semantics and structure of a scene from a single depth image. Our key idea is to deeply fuse an efficient 3D shape estimator with existing recognition (e.g., ResNets) and segmentation (e.g., MaskR-CNN) techniques. Object level semantics and latent feature maps are extracted and then fed to a shape estimator to extract the 3D shape. Extensive experiments are conducted on large-scale synthesized indoor scene datasets, quantitatively and qualitatively demonstrating the merits and superior performance of 3R-Depth.