Deep Learning on Point Clouds and Its Application: A Survey.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Processing

Background:

• Point clouds are essential 3D data formats from sensors like LIDAR and RGB-D cameras.
• The irregular nature of point clouds necessitates specialized feature engineering for deep learning.
• Deep learning excels with structured data like images, prompting adaptation for 3D point clouds.

Purpose of the Study:

• To classify and analyze recent point cloud feature learning methods.
• To provide a comprehensive overview of point cloud applications in 3D computer vision.
• To identify future research directions in point cloud deep learning.

Main Methods:

• Categorization of methods into point-based and tree-based approaches.
• Point-based methods process raw point clouds directly.
• Tree-based methods utilize structures like k-dimensional trees (Kd-trees) for regular representation.

Main Results:

• Analysis of advantages and disadvantages for both point-based and tree-based methods.
• Compilation of datasets and evaluation metrics for point cloud tasks.
• Overview of key applications: 3D object classification, semantic segmentation, and detection.

Conclusions:

• Feature learning is crucial for addressing point cloud irregularity.
• Both point-based and tree-based methods offer distinct advantages for 3D data analysis.
• The field is rapidly evolving with ongoing research in deep learning for point clouds.