Related Experiment Videos
Robust point cloud lightweighting with multi-scale adaptive filtering and entropy-driven subdivision
Weibo Zeng1,2, Xinyu Gao2, Qi Lu2
1Anhui Provincial Key Laboratory of Realistic Geographic Environment, Chuzhou, China.
Plos One
|July 24, 2026
Summary
This study introduces a lightweight point cloud simplification framework that effectively balances noise reduction and geometric detail preservation. The new method significantly improves noise removal, edge retention, and topological accuracy for 3D reconstruction tasks.
Area of Science:
- Computer Vision
- Geometric Modeling
- Data Processing
Background:
- Current point cloud simplification methods struggle to balance noise robustness with geometric detail preservation.
- Complex ground objects, such as buildings and vegetation, present unique challenges for accurate 3D representation.
Purpose of the Study:
- To develop a lightweight point cloud simplification framework that enhances noise suppression while preserving geometric details.
- To improve the accuracy and efficiency of 3D reconstruction and geographic information modeling.
Main Methods:
- A novel framework integrating multi-scale adaptive filtering, entropy-driven spatial partitioning, and an enhanced medial axis transform (MAT).
- Key innovations include an adaptive sliding window filter, a curvature-weighted MAT algorithm, and entropy-driven axis-aligned bounding box (AABB) partitioning.
- The framework was validated on self-collected datasets (buildings, vegetation, roads) and the STPLS3D benchmark.
Main Results:
- Achieved an average noise removal rate of 87.76% (11.99% improvement over baseline) and edge retention of 83.3% (7.35% improvement).
- Skeleton extraction demonstrated superior topological integrity (0.93) and branch accuracy (0.95).
- Lowest mean error in normal estimation (3.34%), and minimal point-to-surface distance and fracture rates in 3D reconstruction.
Conclusions:
- The proposed framework offers a robust solution for point cloud simplification, outperforming existing methods in accuracy and efficiency.
- This method is suitable for applications in 3D geographic information modeling and complex scene reconstruction.
- The integration of adaptive filtering, entropy-driven partitioning, and enhanced MAT provides a significant advancement in point cloud processing.