Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Rethinking Data Input for Point Cloud Upsampling

Tongxu Zhang

TL;DR

This work questions the default patch-based input paradigm for point cloud upsampling by introducing Average Segment (AS), a full-model input. Using the PU-GCN backbone, experiments on PU1K and ABC10K compare Patch and AS inputs across multiple upsampling scales, with CD and HD as primary metrics. Results show Patch-based inputs outperform AS on both datasets, while ablations reveal boundary processing and local feature extraction as key drivers of performance. The study suggests that while AS can retain global context, effective boundary handling and local detail—embodied by patch-based processing—remain superior under the tested conditions, pointing to future work on integrating global and local information through encoding strategies.

Abstract

Point cloud upsampling is crucial for tasks like 3D reconstruction. While existing methods rely on patch-based inputs, and there is no research discussing the differences and principles between point cloud model full input and patch based input. Ergo, we propose a novel approach using whole model inputs i.e. Average Segment input. Our experiments on PU1K and ABC datasets reveal that patch-based inputs consistently outperform whole model inputs. To understand this, we will delve into factors in feature extraction, and network architecture that influence upsampling results.

Rethinking Data Input for Point Cloud Upsampling

TL;DR

This work questions the default patch-based input paradigm for point cloud upsampling by introducing Average Segment (AS), a full-model input. Using the PU-GCN backbone, experiments on PU1K and ABC10K compare Patch and AS inputs across multiple upsampling scales, with CD and HD as primary metrics. Results show Patch-based inputs outperform AS on both datasets, while ablations reveal boundary processing and local feature extraction as key drivers of performance. The study suggests that while AS can retain global context, effective boundary handling and local detail—embodied by patch-based processing—remain superior under the tested conditions, pointing to future work on integrating global and local information through encoding strategies.

Abstract

Point cloud upsampling is crucial for tasks like 3D reconstruction. While existing methods rely on patch-based inputs, and there is no research discussing the differences and principles between point cloud model full input and patch based input. Ergo, we propose a novel approach using whole model inputs i.e. Average Segment input. Our experiments on PU1K and ABC datasets reveal that patch-based inputs consistently outperform whole model inputs. To understand this, we will delve into factors in feature extraction, and network architecture that influence upsampling results.
Paper Structure (14 sections, 9 equations, 5 figures, 4 tables, 1 algorithm)

This paper contains 14 sections, 9 equations, 5 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Methodology
  3. Overview
  4. Data Input Based on Patch
  5. Data Input Based on Average Segment
  6. Experiments
  7. Settings
  8. Evaluation Metrics
  9. Results
  10. Ablations
  11. Visualizations
  12. Discussion
  13. Future Works
  14. Conclusion

Figures (5)

  • Figure 1: The original Mesh models have been transformed into a point cloud through patch based sampling and the Average Segment method proposed in this article.
  • Figure 2: The PU-GCN network structure used in the experiment is shown in the figure, where (a) is the original PU-GCN structure. (b), (c) and (d) are all architectures used in the ablation experiment to explore the modules that have the greatest impact on upsampling.
  • Figure 3: Visualization of PU-GCN with Different Architectures on the PU1K Dataset.
  • Figure 4: Visualization example on the ABC dataset.
  • Figure : Point Cloud Upsampling with Patch Segmentation