Related Experiment Video
Updated: Jul 31, 2025
07:05
Author Spotlight: Innovative Device Development for Advancing Dendroecology and Wood Anatomy Research
Published on: September 27, 2024
2.8K
TGPS: dynamic point cloud down-sampling of the dense point clouds for Terracotta Warrior fragments
Optics Express
|May 9, 2023
Summary
A novel task-driven down-sampling method, TGPS, efficiently processes dense 3D point clouds by learning point importance. This approach enhances downstream tasks like classification and reconstruction, achieving superior accuracy on Terracotta Warrior data.
Area of Science:
- Computer Vision
- 3D Data Processing
- Machine Learning
Background:
- Dense 3D point clouds from 3D scanners contain redundant data, hindering efficient transmission and processing.
- Existing sampling methods generate points that are not learnable or relevant for downstream tasks.
Purpose of the Study:
- To propose an end-to-end, task-driven, and learnable down-sampling method for dense 3D point clouds.
- To address the limitations of current sampling techniques in feature learning and task relevance.
Main Methods:
- Introduced the Task-driven Graph Pooling and Sampling (TGPS) method.
- Utilized point-based Transformer units for feature embedding and global feature extraction.
- Employed Dynamic Graph Attention Edge Convolution (DGA EConv) for local feature aggregation.
- Developed networks for point cloud classification and reconstruction tasks.
Main Results:
- TGPS enables down-sampling guided by global features, retaining points crucial for specific tasks.
- The TGPS-DGA-Net achieved state-of-the-art accuracy in point cloud classification.
- Demonstrated superior performance on both real-world Terracotta Warrior fragments and public datasets.
Conclusions:
- The proposed TGPS method effectively reduces data redundancy while preserving task-relevant information in 3D point clouds.
- TGPS-DGA-Net offers a significant advancement in point cloud classification and reconstruction accuracy.
