TDNet: transformer-based network for point cloud denoising.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Processing

Background:

• Point cloud data is susceptible to noise, degrading its utility in various applications.
• Existing denoising methods often struggle with complex noise patterns and preserving fine geometric details.

Purpose of the Study:

• To propose a novel transformer-based end-to-end network (TDNet) for effective point cloud denoising.
• To leverage natural language processing transformer architectures for enhanced point cloud feature extraction.

Main Methods:

• Developed TDNet, an encoder-decoder network utilizing a transformer architecture adapted for point clouds.
• Employed an adaptive sampling strategy to reconstruct surfaces based on latent manifold learning.
• Treated points as 'words' to adapt transformer's self-attention mechanism for semantic relevance in point clouds.

Main Results:

• TDNet achieved superior quantitative and qualitative denoising results on synthetic datasets.
• The network demonstrated effectiveness in processing real-world noisy point cloud data, including terracotta warrior fragments.
• The proposed method successfully extracts features and reconstructs clean point clouds.

Conclusions:

• TDNet offers a powerful new approach to point cloud denoising by adapting transformer architectures.
• The adaptive sampling method enhances surface reconstruction accuracy.
• The study validates the effectiveness of TDNet for both synthetic and real-world noisy point cloud data.