PatchNets: Patch-Based Generalizable Deep Implicit 3D Shape Representations

Abstract

In this supplemental material, we expand on some points from the main paper. We first perform an ablation study on the extrinsics losses in Sec. 1. In Sec. 2, we describe the error measures we employ. In Sec. 3, we compare error measures on the reduced and full test sets. Sec. 4 shows the randomly picked single shape that we use in one of the generalization experiments. Sec. 5 contains more experiments using object-level priors. Sec. 6 shows different number of patches and network/latent code sizes. Next, we measure the performance under synthetic noise in Sec. 7. We show preliminary results on a large scene in Sec. 8. Finally, in Sec. 9, we provide some remarks on the concurrent work DSIF.