How matterport/mask_rcnn works
This repository is a Keras/TensorFlow Mask R-CNN implementation for object detection and instance segmentation, with examples for inference, training, and evaluation. Users provide images to a configured model, which returns per-instance boxes, classes, scores, and masks; visualization helpers can display those results. Training examples adapt datasets and configurations, while the COCO example also supports AP evaluation. The README documents the detection stages, but excerpts do not expose their complete internal call graph, so stage links cite the README; unsampled sample implementations are represented only at their supported directory paths.
Main entry points: model.py Mask R-CNN model · utils.py Dataset interface · coco.py COCO train/evaluate · config.py Model configuration
Components 17
Detection model
- Mask R-CNN model
mrcnn/model.py - Region proposals
mrcnn/model.py - Box classification
mrcnn/model.py - Instance masks
mrcnn/model.py - Detection results
mrcnn/model.py
Data and configuration
- Model configuration
mrcnn/config.py - Dataset interface
mrcnn/utils.py - Image preparation
mrcnn/utils.py
Examples and evaluation
- COCO train/evaluate
samples/coco/coco.py - Custom datasets
samples - Balloon example
samples/balloon/balloon.py - Nucleus example
samples/nucleus/nucleus.py - Shapes example
samples/shapes/shapes.py - COCO AP metrics
samples/coco/coco.py
Visualization and parallelism
- Result visualization
mrcnn/visualize.py - Multi-GPU training
mrcnn/parallel_model.py
Other components
- Image user
Connections 18
- Image user → Image preparation: provides images
README.md - Image preparation → Mask R-CNN model: prepares inputs
mrcnn/model.py - Model configuration → Mask R-CNN model: configures
mrcnn/model.py - Mask R-CNN model → Region proposals: generates proposals
README.md - Region proposals → Box classification: refines boxes
README.md - Box classification → Instance masks: predicts masks
README.md - Mask R-CNN model → Detection results: returns detections
mrcnn/model.py - Detection results → Result visualization: renders results
mrcnn/visualize.py - Mask R-CNN model → Dataset interface: uses utilities
mrcnn/model.py - Dataset interface → Image preparation: supports image prep
README.md - COCO train/evaluate → Mask R-CNN model: trains and evaluates
README.md - COCO train/evaluate → Dataset interface: adapts dataset
README.md - COCO train/evaluate → COCO AP metrics: reports AP
README.md - Custom datasets → Dataset interface: extends interface
README.md - Balloon example → Dataset interface: demonstrates adaptation
README.md - Nucleus example → Dataset interface: demonstrates adaptation
README.md - Shapes example → Dataset interface: demonstrates adaptation
README.md - Multi-GPU training → Mask R-CNN model: replicates model
mrcnn/model.py