charlesq34/pointnet
Pointnet
PointNet ist eine Deep-Learning-Architektur, die darauf ausgelegt ist, rohe 3D-Punktwolken direkt ohne Voxelisierung zu verarbeiten und zu klassifizieren. Es bietet ein System für die 3D-Objektklassifizierung, Frameworks für die semantische Segmentierung zur Unterteilung von Wolken in Kategorien sowie Tools zur Visualisierung von 3D-Formen.
Das Projekt nutzt ein Transform-Netzwerk, um Punktwolken in einen kanonischen Koordinatenraum auszurichten, und verwendet eine auf symmetrischen Funktionen basierende Aggregation, um punktweise Merkmale unabhängig von der Punktreihenfolge in globale Vektoren zu verdichten. Zudem verfügt es über eine Multi-Scale-Grouping-Architektur, um hierarchische geometrische Merkmale über verschiedene räumliche Skalen hinweg zu extrahieren.
Das System umfasst Datenpipelines zur Konvertierung von rohen 3D-Innenraum-Parsing- und Formdaten in HDF5-Dateien für das Training. Die Evaluierungsfunktionen decken Genauigkeitsmetriken für die Segmentierung sowie das Rendering von falsch klassifizierten 3D-Punktwolken in Drei-Ansichten-Bilder zur Fehleranalyse ab.
Features
- Point Cloud Grouping - Employs multi-scale grouping techniques to extract hierarchical geometric features from raw 3D point clouds.
- Point Cloud Neural Architectures - Provides a neural network architecture designed to process and classify raw 3D point clouds directly without voxelization.
- Symmetric Aggregation Functions - Uses a max-pooling symmetric function to condense point-wise features into a single global vector regardless of point order.
- Point Cloud and 3D Processing - Processes point sets sampled from a surface to assign a single categorical label to the entire 3D shape.
- Point-Wise Label Prediction - Assigns semantic categories to every single point in a point cloud using shared weights.
- Semantic Segmentations - Provides a framework for assigning semantic labels to individual points within 3D indoor scenes.
- Semantic Part Segmentation - Partitions 3D point clouds into constituent parts using trained models to predict point-wise labels.
- Neural Spatial Transformers - Features a T-Net transform network that learns a spatial transformation matrix to align point clouds into a canonical space.
- 3D Point Cloud Custom Training - Converts raw 3D point cloud data into structured HDF5 formats compatible with deep learning training pipelines.
- Global-Local Feature Fusion - Combines global shape descriptors with local point features to enable precise semantic labeling of individual coordinates.
- 3D Data Pipeline Workflows - Implements a data processing workflow that converts raw 3D indoor parsing and shape data into HDF5 files.
- 3D Indoor Scene Dataset Exporters - Provides workflows for converting raw 3D indoor scene data into structured HDF5 files for model training.
- Training Data Serialization - Implements an HDF5-based serialization system for efficient disk access to large-scale 3D point clouds during training.
- Point Cloud Visualizers - Includes utilities for rendering raw 3D shapes and generating three-view images for error analysis.
- 3D Computer Vision - Deep learning architecture for 3D classification and segmentation.
- 3D Object Classification - Deep learning on point sets for classification and segmentation.
- Deep Learning Architectures - Foundational deep learning architecture for point set classification and segmentation.
- Deep Learning Frameworks - Pioneering deep learning architecture for point cloud classification and segmentation.
- Semantic Segmentation - Deep learning framework for classification and segmentation on point sets.
