Self-Sampling for Neural Point Cloud Consolidation

We introduce a novel technique for neural point cloud consolidation which learns from only the input cloud. Unlike other up-sampling methods analyze shapes via local patches, in this work, we learn global subsets. repeatedly self-sample with subsets that are used to train deep network. Specifically, define source and target according desired criteria (e.g., generating sharp points or sparse regions). The network mapping subsets, implicitly consolidate During inference, is fed random of input, it displaces synthesize consolidated set. leverage inductive bias networks eliminate noise outliers, notoriously difficult problem consolidation. shared weights optimized over entire shape, learning non-local statistics exploiting recurrence local-scale geometries. encodes distribution underlying shape surface within fixed set kernels, results best explanation surface. demonstrate ability sets variety shapes, while eliminating outliers noise.