Reproduce “Self-Supervised Implicit Attention Priors for Point Cloud Reconstruction” on OpenResearch

About this paper

Kyle Fogarty, Chenyue Cai, Jing Yang, Zhilin Guo, Cengiz Öztireli

The authors reconstruct surfaces from sparse, irregular point clouds using a shape-specific implicit field whose query locations attend to a shared trainable dictionary, then refine the field with RIMLS using sampled points and gradient-derived normals. On a curated self-similar-object dataset, the method reports CD 0.017, NC 0.907, and F-score 72.44, but PG-SDF has lower Hausdorff Distance.

Key insights

• Non-local self-similarity acts as a potent geometric prior, which can be explicitly learned from a single input point cloud through a cross-attention mechanism with a learnable dictionary.
• A dictionary-conditioned neural field, integrating positional encoding and multi-headed attention, effectively captures shape-specific recurring patterns and long-range correlations within an implicit representation.
• A hybrid approach, combining a neural field's global guidance for densification and normal estimation with Robust Implicit Moving Least Squares (RIMLS) for local detail preservation, yields high-fidelity and robust surface reconstructions.

Summary by alphaXiv.