PU-DZMS: Point Cloud Upsampling via Dense Zoom Encoder and Multi-Scale Complementary Regression.

Area of Science:

• Computer Vision
• 3D Geometry Processing
• Machine Learning

Background:

• Point cloud sparsity in imaging leads to loss of critical geometric detail.
• Existing point cloud upsampling networks struggle with local-global feature understanding, causing contour distortion and sparse regions.

Purpose of the Study:

• To address limitations in current point cloud upsampling techniques.
• To propose a novel method, PU-DZMS, for enhanced point cloud density and detail recovery.

Main Methods:

• The proposed PU-DZMS method comprises two key components: the Dense Zoom Encoder (DENZE) and the Multi-Scale Complementary Regression (MSCR) module.
• DENZE utilizes ZOOM Blocks with dense connections and a Transformer mechanism to capture local-global geometric features, clarifying point cloud edges.
• MSCR expands features and regresses dense point clouds using cross-scale residual learning, ensuring geometric continuity and reducing local sparsity.

Main Results:

• Experimental results on the PU-GAN and PU-Net datasets demonstrate the effectiveness of PU-DZMS.
• The method successfully enhances geometric detail and reduces sparse regions in point clouds.
• PU-DZMS shows strong performance in point cloud upsampling tasks.

Conclusions:

• PU-DZMS effectively overcomes limitations of existing methods in local-global relation understanding for point cloud upsampling.
• The proposed architecture clarifies geometric edges and reduces local sparse regions, leading to improved point cloud quality.