LassoNet: Deep Lasso-Selection of 3D Point Clouds.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Analysis

Background:

• 3D point cloud selection is crucial for exploratory analysis but current heuristic methods struggle with data variability.
• Challenges include variations in point cloud density, occlusions, and lasso selection size.

Purpose of the Study:

• Introduce LassoNet, a deep neural network for robust lasso selection in 3D point clouds.
• Learn a latent mapping from user input (viewpoint, lasso) to desired point cloud regions.
• Improve scalability and effectiveness of point cloud selection methods.

Main Methods:

• Developed LassoNet, a hierarchical deep neural network trained on over 30,000 lasso selection records.
• Coupled user-target points with viewpoint and lasso data using 3D coordinate transforms.
• Enhanced scalability through intention filtering and farthest point sampling.

Main Results:

• LassoNet demonstrated improved selection effectiveness and efficiency compared to state-of-the-art methods.
• Performance gains were consistent across various 3D point cloud types, viewpoints, and lasso selections.
• A formal user study validated the superiority of the LassoNet approach.

Conclusions:

• LassoNet offers a significant advancement in automated 3D point cloud selection.
• The deep learning approach overcomes limitations of traditional heuristic methods.
• This work provides a more effective and efficient tool for 3D data exploration.