https://github.com/carlostojal/ndnetpp

ND-Net++ is a PointNet++-based point cloud classification, part segmentation and semantic segmentation neural network. ND is an abbreviation for "Normal Distribution".
https://github.com/carlostojal/ndnetpp

Last synced: about 1 month ago
JSON representation

ND-Net++ is a PointNet++-based point cloud classification, part segmentation and semantic segmentation neural network. ND is an abbreviation for "Normal Distribution".

• Host: GitHub
• URL: https://github.com/carlostojal/ndnetpp
• Owner: carlostojal
• License: mit
• Created: 2024-08-02T23:42:30.000Z (4 months ago)
• Default Branch: main
• Last Pushed: 2024-08-29T22:10:39.000Z (3 months ago)
• Last Synced: 2024-08-30T12:46:39.079Z (3 months ago)
• Language: Python
• Homepage:
• Size: 76.2 KB
• Stars: 1
• Watchers: 2
• Forks: 0
• Open Issues: 4
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# ND-Net++

Created by Carlos Tojal, Daniel Carias, Luis Conde Bento, Hugo Costelha and Catarina Reis from Polytechnic Institute of Leiria.

ND-Net++ is a PointNet++-based point cloud classification, part segmentation and semantic segmentation neural network. ND is an abbreviation for "Normal Distribution".

PointNet++ adopts multiple sampling and grouping stages to hierarchically extract features while abstracting the point cloud. To reduce the size of the network, we propose a new architecture sharing the same hierarchical concept, but estimating normal distributions instead of performing sampling and grouping.

The grouping stage of the PointNet++ generates $K$ groups with $N$ 3-dimensional neighbors ($B \times K \times N \times 3$ tensor). Instead, a normal distribution can represent a neighborhood of points by its mean 3-dimensional vector and its $3 \times 3$ covariance matrix only.