3D Shape Reconstruction of Furniture Object from a Single Real Indoor Image

Most of the current researches on 3D shape reconstruction based on deep learning only focus on clean-background images. In this paper, we propose a system that segments furniture object from a single real indoor image and reconstructs their 3D meshes. This system takes Pixel2Mesh, a 3D shape reconstruction network, as a branch of the state-of-the-art instance segmentation network Mask Scoring R-CNN. We trained our system on 3D-FUTURE dataset, and the results show that our method can effectively reconstruct 3D shape of furniture objects from real indoor images by designing proper loss functions and combining instance segmentation and 3D shape reconstruction network.