How charlesq34/pointnet works
Main Components Identification
The system should be divided into three main functional blocks:
• Classification Pipeline (/models/pointnet_cls.py)
• Part Segmentation Pipeline (/part_seg/)
• Semantic Segmentation Pipeline (/sem_seg/)
Core Architecture Elements
a) Input Processing Layer:
• Point cloud input (N x 3 points)
• Transformation network (transform_nets.py)
• Input normalization process
b) Feature Extraction:
• Shared MLP layers
• Max pooling operations
• Feature transformation components
c) Task-Specific Heads:
• Classification branch
• Part segmentation branch
• Semantic segmentation branch
[Input Layer] → [Transform Net] → [Feature Extraction] → [Global Feature Aggregation]
• ↘
• [Task-Specific Heads]
Important Details
a) Data Pipeline:
• Data preprocessing steps (provider.py)
• Utility functions (utils/)
• Demonstrate the training pipeline flow
b) Model Architecture:
• Permutation invariance property
• Global and local feature learning
• Point feature transformation
Specific Implementation Notes
• Classification network (pointnet_cls.py)
• Part segmentation (part_seg/pointnet_part_seg.py)
• Semantic segmentation (sem_seg/model.py)
Technical Specifications
• Input: N×3 point cloud
• Feature dimensions at each layer
• Output dimensions for each task
• Key hyperparameters
Integration Points
Connections with:
• Data preparation utilities (utils/data_prep_util.py)
• Evaluation metrics (evaluate.py)
• Training procedures (train.py)