Source-linked AI summary
SnowflakeNet: Point Cloud Completion by Snowflake Point Deconvolution with Skip-Transformer
Peng Xiang, Xin Wen, Yu-Shen Liu, Yan-Pei Cao, Pengfei Wan, Wen Zheng, Zhizhong Han
TL;DR
Point cloud completion must infer detailed complete shapes from partial observations despite discrete points and unstructured local prediction. SnowflakeNet progressively splits parent points with Snowflake Point Deconvolution and uses a skip-transformer to transfer local splitting patterns across levels. Experiments report state-of-the-art performance on widely used completion benchmarks.
Problem
Point cloud completion is limited by discrete point representations and unstructured local predictions, making fine geometric details difficult to reveal from partial observations.
Method
SnowflakeNet progressively generates child points by splitting parent points with SPD, while skip-transformer attention transfers previous-layer splitting patterns to current-layer decoding.
Results
SnowflakeNet achieves state-of-the-art completion accuracy and outperforms current state-of-the-art methods on sparse and dense benchmark datasets.
Takeaways & Limitations
The method produces locally compact, structured point clouds that capture detailed geometry such as smooth regions, sharp edges, and corners.
Abstract
from arXiv · showhide
Point cloud completion aims to predict a complete shape in high accuracy from its partial observation. However, previous methods usually suffered from discrete nature of point cloud and unstructured prediction of points in local regions, which makes it hard to reveal fine local geometric details on the complete shape. To resolve this issue, we propose SnowflakeNet with Snowflake Point Deconvolution (SPD) to generate the complete point clouds. The SnowflakeNet models the generation of complete point clouds as the snowflake-like growth of points in 3D space, where the child points are progressively generated by splitting their parent points after each SPD. Our insight of revealing detailed geometry is to introduce skip-transformer in SPD to learn point splitting patterns which can fit local regions the best. Skip-transformer leverages attention mechanism to summarize the splitting patterns used in the previous SPD layer to produce the splitting in the current SPD layer. The locally compact and structured point cloud generated by SPD is able to precisely capture the structure characteristic of 3D shape in local patches, which enables the network to predict highly detailed geometries, such as smooth regions, sharp edges and corners. Our experimental results outperform the state-of-the-art point cloud completion methods under widely used benchmarks. Code will be available at https://github.com/AllenXiangX/SnowflakeNet.
1. Introduction
SnowflakeNet addresses incomplete point-cloud completion by progressively splitting points and using cross-level attention to preserve local geometric structure. The resulting locally structured generation targets detailed surfaces, including smooth regions and sharp edges.
- Motivation: Point cloud completion predicts a complete shape from sparse, incomplete observations affected by occlusion and limited sensor resolution.Existing methods often rely on hierarchical rooted trees or assumed target topologies, while suffering from discrete and locally unstructured point generation.
- SnowflakeNet: SnowflakeNet models completion as snowflake-like growth, progressively generating child points by splitting parent points across stacked SPD layers.Each child inherits shape characteristics captured by its parent point.
- Skip-Transformer: Skip-transformer learns local point-splitting patterns by using attention to summarize the previous SPD layer and guide splitting in the current layer.It integrates spatial relationships across decoding levels rather than modeling attention only within one level.
- Geometric Detail: The method is designed to capture local 3D surface characteristics and predict detailed geometries across smooth planes, sharp edges, and corners.The skip-transformer uses local shape context and spatial relationships between parent and child points to refine these structures.
- Contribution: SnowflakeNet explicitly represents complete-shape generation as a locally structured pattern rather than locally unorganized point prediction.The contribution list characterizes this structure as improving 3D shape completion performance.
2. Related Work
Prior point-cloud completion methods use learned priors, folding or coarse-to-fine decoding, and explicit topologies, but often lack locally structured generation. SnowflakeNet combines an explicit hierarchical process with locally organized point splitting to target detailed geometry.
- Method Categories: Point cloud completion methods include traditional approaches and deep-learning methods that infer missing regions or complete shapes from learned priors.SnowflakeNet belongs to the deep-learning category and focuses specifically on the decoding process.
- Folding-Based Decoding: Folding-based methods increase point density through feature expansion, but shape information represented only by point features is not explicitly constrained or explained.The related-work discussion identifies this as a limitation of folding-based decoding.
- Coarse-to-Fine Decoding: Coarse-to-fine methods first generate a low-resolution cloud and then use a lifting module to increase point density.PCN and NSFA are cited as examples of this two-stage generation framework.
- SnowflakeNet: SnowflakeNet models progressive completion as an explicit hierarchical rooted-tree process while preserving an explainable and explicit generation procedure.The approach is positioned as combining properties associated with TopNet, CDN, and PF-Net.
- Local Structure: Unlike earlier approaches, SnowflakeNet organizes point splitting locally, supporting precise capture of detailed 3D-shape geometries and structures.This related-work comparison links local organization with detailed geometric representation.
3. SnowflakeNet
SnowflakeNet completes point clouds through staged Snowflake Point Deconvolution, progressively splitting parent points into locally structured child points. Its skip-transformer links displacement information across decoding levels to refine local geometric detail, while training combines completion and preservation objectives.
- Overview: SnowflakeNet extracts global and local shape information, generates a coarse complete seed cloud, and then applies three point-generation modules.The architecture contains feature extraction, seed generation, and point generation; the feature extractor uses set abstraction and point transformers.
- Snowflake Point Deconvolution: Point-wise splitting combines activated shape characteristics from parent features with shared learnable kernels to generate multiple child point features.Kernel logits represent activation status, while child features are formed by adding the activated characteristics; ri = 1 can reposition points, whereas ri > 1 expands them.
- Snowflake Point Deconvolution: Each SPD duplicates parent points, predicts displacements, and updates the duplicated points to produce child points at progressively higher densities.For parent points Pi−1, duplication by factor ri forms ˆPi, and the output is Pi = ˆPi + ∆Pi.
- Skip-Transformer: The skip-transformer integrates current point features with previous displacement features and uses local k-nearest-neighbor attention to refine cross-level spatial context.The skip connection carries displacement information from the previous layer, and attention weights determine how much the current splitting uses previous splitting information.
- Training Loss: Training minimizes four density-matched Chamfer-distance completion losses together with a partial matching preservation loss.Ground-truth clouds are down-sampled to match Pc, P1, P2, and P3 before the completion losses are summed.
4. Experiments
Experiments on PCN and Completion3D show that SnowflakeNet outperforms prior completion methods quantitatively and produces smoother, more detailed reconstructions visually. Ablations further support the contributions of skip-transformer, point-wise splitting, and the overall architecture.
- PCN dataset: 0.85 lower average CD, or 10.5% below NSFA, is reported on PCN using L1 Chamfer distance.The comparison uses the PCN evaluation protocol and identifies NSFA as the second-ranked method.
- Benchmark evaluation: SnowflakeNet achieves the best reported performance on both PCN and Completion3D benchmarks.On PCN it outperforms all counterparts; on Completion3D it achieves the best results over all leaderboard methods.
- Visual comparison: Visual comparisons show smoother surfaces and more detailed structures, including car boundaries, chair backs, plane surfaces, boat sails, and table legs.These qualitative improvements are reported on both PCN and Completion3D examples.
- Completion3D dataset: 1.63 lower average CD, or 17.3% below PMP-Net, is reported on Completion3D using L2 Chamfer distance.SnowflakeNet also outperforms other listed methods in every Completion3D category.
- Ablation studies: The Full transformer-based model performs best among skip-transformer variants, supporting cross-layer connections and transformer attention for learning local shape context.Comparisons with No-connect, No-att, and Self-att variants isolate these design choices.
- Point splitting behavior: Snowflake point deconvolution generates child points around parent points while preserving smooth planes, thin tubes, and sharp-edge geometry.Visualization of two successive splitting layers illustrates the progressive point-splitting paths.
5. Conclusions
SnowflakeNet completes partial point clouds through snowflake-like point growth, using SPD and skip-transformer to produce locally compact, structured outputs with detailed geometries. Experiments on sparse and dense completion datasets show superiority over current state-of-the-art methods.
- SnowflakeNet models completion as snowflake-like growth in 3D space using multiple Snowflake Point Deconvolution layers.Each SPD layer progressively generates points by splitting parent points.
- Skip-transformer enables SnowflakeNet to generate locally compact and structured point clouds with highly detailed geometries.
- Experiments on Completion3D and PCN show SnowflakeNet’s superiority over current state-of-the-art point cloud completion methods.The evaluation covers sparse and dense point cloud completion datasets.