graph LR
TF_Semantic_Segmentation_Pipeline["TF Semantic Segmentation Pipeline"]
PyTorch_Semantic_Segmentation_Pipeline["PyTorch Semantic Segmentation Pipeline"]
TF_Object_Detection_Pipeline["TF Object Detection Pipeline"]
PyTorch_Object_Detection_Pipeline["PyTorch Object Detection Pipeline"]
Data_Management_Layer["Data Management Layer"]
Model_Architectures["Model Architectures"]
Configuration_and_Checkpointing["Configuration and Checkpointing"]
Metrics_and_Visualization["Metrics and Visualization"]
TF_Semantic_Segmentation_Pipeline -- "Retrieves data from" --> Data_Management_Layer
TF_Semantic_Segmentation_Pipeline -- "Interacts with" --> Model_Architectures
TF_Semantic_Segmentation_Pipeline -- "Manages" --> Configuration_and_Checkpointing
TF_Semantic_Segmentation_Pipeline -- "Utilizes" --> Metrics_and_Visualization
PyTorch_Semantic_Segmentation_Pipeline -- "Retrieves data from" --> Data_Management_Layer
PyTorch_Semantic_Segmentation_Pipeline -- "Interacts with" --> Model_Architectures
PyTorch_Semantic_Segmentation_Pipeline -- "Manages" --> Configuration_and_Checkpointing
PyTorch_Semantic_Segmentation_Pipeline -- "Utilizes" --> Metrics_and_Visualization
TF_Object_Detection_Pipeline -- "Retrieves data from" --> Data_Management_Layer
TF_Object_Detection_Pipeline -- "Interacts with" --> Model_Architectures
TF_Object_Detection_Pipeline -- "Manages" --> Configuration_and_Checkpointing
TF_Object_Detection_Pipeline -- "Utilizes" --> Metrics_and_Visualization
PyTorch_Object_Detection_Pipeline -- "Retrieves data from" --> Data_Management_Layer
PyTorch_Object_Detection_Pipeline -- "Interacts with" --> Model_Architectures
PyTorch_Object_Detection_Pipeline -- "Manages" --> Configuration_and_Checkpointing
PyTorch_Object_Detection_Pipeline -- "Utilizes" --> Metrics_and_Visualization
The ml3d project's core functionality is orchestrated by specialized Pipeline components for TensorFlow and PyTorch, handling both semantic segmentation and object detection tasks. These pipelines serve as the central orchestrators, managing the entire machine learning lifecycle from data ingestion to model evaluation. They retrieve data from the Data Management Layer, which is responsible for efficient loading, preprocessing, and augmentation of 3D point cloud datasets. The pipelines interact with the Model Architectures component, which encapsulates the various deep learning models tailored for specific tasks and frameworks. Throughout the training and inference processes, the pipelines manage the Configuration and Checkpointing component to apply experiment settings and persist model states. Finally, they utilize the Metrics and Visualization component to compute performance metrics and generate insightful 3D visualizations of the results. This modular design ensures clear separation of concerns, allowing for flexible integration of new datasets, models, and evaluation techniques.
Manages the sequential execution of machine learning stages (training, validation, testing, inference) for TensorFlow semantic segmentation models. It drives forward/backward passes, coordinates data loading, handles model checkpointing, records metrics, applies configurations, and produces summaries including 3D visualizations.
Related Classes/Methods:
Manages the sequential execution of machine learning stages (training, validation, testing, inference) for PyTorch semantic segmentation models. It drives forward/backward passes, coordinates data loading, handles model checkpointing, records metrics, applies configurations, and produces summaries including 3D visualizations.
Related Classes/Methods:
Manages the sequential execution of machine learning stages (training, validation, testing, inference) for TensorFlow object detection models. It drives forward/backward passes, coordinates data loading, handles model checkpointing, records metrics, applies configurations, and produces summaries including 3D visualizations.
Related Classes/Methods:
Manages the sequential execution of machine learning stages (training, validation, testing, inference) for PyTorch object detection models. It drives forward/backward passes, coordinates data loading, handles model checkpointing, records metrics, applies configurations, and produces summaries including 3D visualizations.
Related Classes/Methods:
Responsible for loading, preprocessing, and augmenting 3D point cloud data from various datasets, providing a unified interface for different machine learning pipelines. It handles data iterators and batching for efficient model training and evaluation.
Related Classes/Methods:
Contains the implementations of various 3D deep learning models for tasks like semantic segmentation and object detection, supporting both TensorFlow and PyTorch frameworks. It defines the neural network structures and their forward pass logic.
Related Classes/Methods:
Manages the loading and application of experiment configurations, and handles the saving and loading of model checkpoints during training and inference. This ensures reproducibility and allows for resuming training or deploying trained models.
Related Classes/Methods:
Provides functionalities for calculating and reporting performance metrics (e.g., mAP, IoU) and generating visualizations, including 3D point cloud visualizations, to assess model performance and provide insights.
Related Classes/Methods: