PU-Dense: Sparse Tensor-Based Point Cloud Geometry Upsampling.

Area of Science:

• Computer Vision
• 3D Geometry Processing
• Machine Learning

Background:

• Augmented and virtual reality drive demand for high-resolution point clouds.
• Point cloud data capture and processing pipelines can introduce geometric detail loss and irregularities.
• Existing point cloud upsampling methods struggle with dense, real-world datasets.

Purpose of the Study:

• To develop a novel and efficient point cloud upsampling technique capable of reconstructing high Level-of-Detail (LoD) geometry.
• To address the limitations of current upsampling methods, particularly for dense, photo-realistic, and real-world point clouds.
• To create a robust method that handles diverse point cloud types, including synthetic, LiDAR, and scanned data.

Main Methods:

• PU-Dense utilizes a 3D multiscale architecture with sparse convolutional networks.
• A UNet-like structure with a novel Feature Extraction Unit enables progressive rescaling and multiscale feature learning.
• The method employs a binary voxel occupancy classification loss for memory-efficient processing of high-resolution point clouds.

Main Results:

• PU-Dense significantly outperforms state-of-the-art methods in qualitative and quantitative evaluations.
• The technique demonstrates substantially lower inference time complexity compared to existing approaches.
• The method shows robustness in handling noisy data and is memory efficient for dense point clouds.

Conclusions:

• PU-Dense offers a superior solution for point cloud geometry upsampling, particularly for demanding real-world applications.
• The novel architecture and loss function enable efficient and accurate reconstruction of high-fidelity point cloud data.
• This advancement facilitates improved detail capture for augmented and virtual reality experiences.