Silhouette-assisted 3D object instance reconstruction from a cluttered scene
The objective of our work is to reconstruct 3D object instances from a single RGB image of a cluttered scene. 3D object instance reconstruction is an ill-posed problem due to the presence of heavily occluded and truncated objects, and self-occlusions that lead to substantial regions of unseen areas. Previous works for 3D reconstruction take clues from object silhouettes to carve reconstructed outputs. In this paper, we explore two ways to include silhouette learnable in the network for 3D instance reconstruction from a single cluttered scene image. To this end, in the first approach, we automatically generate instance-specific silhouettes that are compactly encoded within our network design and used to improve the reconstructed 3D shapes; in the second approach, we find an efficient design to regularize object reconstruction explicitly. Experimental results on the SUNCG dataset show that our methods have better performance than the state-of-the-art.