Neural Part Priors: Learning to Optimize Part-Based Object Completion in RGB-D Scans

Figure 1 . Our Neural Part Priors learn optimizable parametric latent spaces of object part geometries, which we can use to fit partial, real-world RGB-D scans of a scene, decomposing detected objects into their complete part geometries. In contrast to 3D scene understanding approaches that make independent predictions per-object, our parametric part spaces enables formulating test-time constraints for consistency within an input scene, thus producing both accurate as well as globally-consistent part decompositions.