How xinge008/cylinder3d works
Main Components (Top Level)
a. Data Processing Layer
b. Model Architecture Layer
c. Training/Inference Layer
Data Flow Components
a. Input Sources (Top Left):
• SemanticKITTI
• nuScenes
• To a "Data Loading" component
b. Data Processing Pipeline:
• PC_Dataset as base class
• Dataset-specific loaders as sub-components
• "Data Builder" component (from builder/data_builder.py)
• Data transformation to cylindrical representation
Model Architecture (Center)
a. Feature Generation:
• It to "Asymmetrical 3D Spconv" component
b. Network Components (from network/ directory):
• Cylinder_spconv_3d module
• Segmentator_3d_asymm_spconv module
Training System (Right Side)
Create components for:
• Loss Function (from builder/loss_builder.py)
• Model Builder (from builder/model_builder.py)
• Training Loop
• Metrics Computation (from utils/metric_util.py)
Configuration System (Top)
Configuration management:
• Dataset configs
• Model parameters
Specific technical details
• PyTorch requirement
• Spconv integration
• Torch-scatter dependency
• Data format transformations
Important notes
• Real-time processing capabilities (10Hz)
• Multi-scan support
• Validation/testing paths
Key Considerations:
• Keep the cylindrical representation as a central element
• Asymmetrical 3D convolution architecture
• Clear data transformation stages
• Performance metrics placement
• Component types
• Processing stages
• Configuration elements
• External dependencies
Level of Detail:
• Detailed network architecture
• Data preprocessing pipeline
• Training workflow