yanx27/Pointnet_Pointnet2_pytorch
Pointnet Pointnet2 Pytorch
This project is a PyTorch-based framework of deep learning models designed for the classification and semantic segmentation of 3D point cloud data. It provides implementations of the PointNet architecture to perform global category labeling of entire objects and detailed partitioning of large-scale 3D environments.
The system handles semantic segmentation across multiple scales, ranging from identifying individual components within a single object to labeling distinct category types within large-scale scenes.
The framework includes structural components for processing unordered point sets, such as farthest-point sampling, hierarchical point-set grouping, and symmetric set-aggregation networks. It also provides offline data preprocessing utilities and tools for generating 3D object files to visualize segmentation outputs in external software.
Features
- 3D Detection and Segmentation - Provides a complete framework for identifying and segmenting objects within large-scale 3D point clouds and scenes.
- PointNet Architectures - Implements the PointNet architecture for capturing local and global features in 3D point clouds.
- Point Cloud Classifications - Implements global category labeling to classify entire 3D objects from point cloud data.
- Point Cloud Neural Architectures - Provides PyTorch-based neural architectures that operate directly on unordered 3D point sets.
- Point-Wise Label Prediction - Implements neural network layers that assign semantic category labels to every individual point in a 3D cloud.
- Global Point Cloud Classifications - Analyzes 3D point sets to assign a single global category label to an entire object.
- Semantic Segmentations - Partitions 3D point clouds into distinct regions by assigning semantic labels to individual points.
- Point Cloud Grouping - Provides multi-scale grouping techniques for hierarchical feature learning in point cloud architectures.
- Point Cloud Segmentation Executions - Identifies and labels individual components within a single 3D object using per-point semantic prediction.
- Large-Scale Scene Segmentations - Labels points in large-scale 3D environments to distinguish between different object types.
- Symmetric Aggregation Functions - Uses shared MLPs and max-pooling to aggregate point-wise features into permutation-invariant global descriptors.
- Feature Interpolations - Calculates weighted averages of nearest neighbors to estimate features for specific query points.
- Point Cloud Processing - Provides offline data transformation and uniform sampling to optimize 3D point cloud training efficiency.
- Sampling - Implements furthest point sampling to select representative subsets of 3D point clouds.
