---

name: mask-rcnn description: "Guides Matterport Mask_RCNN package workflows for instance segmentation, dataset preparation, training, inference, visualization, and evaluation." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

Mask_RCNN Repo Skill

Use this skill when a task names Mask_RCNN, Matterport Mask R-CNN, the mrcnn Python package, mask-rcnn distribution, or asks for Mask R-CNN instance segmentation workflows with Keras/TensorFlow.

This is a self-contained operating guide for the package API and sample workflows. It does not require the original repository checkout; scripts and references here replace the reusable parts of the source samples.

Quick compatibility check

Mask_RCNN is legacy TensorFlow/Keras code. Prefer a Python 3.7-era stack for faithful execution:

python -m pip install "tensorflow==1.15.5" "Keras==2.3.1" "numpy==1.18.5" \
  "h5py==2.10.0" "scikit-image==0.16.2" "opencv-python==4.5.5.64" \
  imgaug pycocotools
python -m pip install mask-rcnn
python - <<'PY'
import mrcnn
from mrcnn.config import Config
from mrcnn import model as modellib
print("mrcnn import ok", mrcnn.__file__)
print("Config", Config.__name__)
print("MaskRCNN", modellib.MaskRCNN.__name__)
PY

If a modern TensorFlow/Keras environment fails with No module named 'keras.engine', tf.log missing, or reshape errors during graph construction, read installation-and-compatibility.md and troubleshooting.md before patching or porting.

For a reusable runtime diagnostic, run:

python scripts/check_env.py --show-signatures

Route by task

• Core package APIs and compatibility: read core-apis for Config, Dataset, MaskRCNN, utility functions, signatures, graph-building constraints, and environment diagnostics.
• Dataset preparation and mask formats: read data-preparation for utils.Dataset subclassing, load_image(), load_mask(), prepare(), VIA/COCO/nucleus layouts, Shapes fixtures, resize modes, and mini-masks.
• Training and fine-tuning: read training for MaskRCNN(mode="training"), pretrained weights, layer selection (heads, 4+, all), schedules, logs/checkpoints, augmentation, and GPU cautions.
• Inference, visualization, and evaluation: read inference-evaluation for MaskRCNN(mode="inference"), detect() outputs, display helpers, color splash, COCO AP conversion, and nucleus RLE output.

Common workflow skeleton

1. Install a compatible TensorFlow/Keras stack and verify imports with scripts/check_env.py.
2. Create a Config subclass with NAME, NUM_CLASSES, image resize dimensions, GPU_COUNT, and IMAGES_PER_GPU appropriate to the workflow.
3. For training, create dataset subclasses derived from mrcnn.utils.Dataset; implement load_*, load_mask, and optionally image_reference; call prepare() before training.
4. Create modellib.MaskRCNN(mode="training"|"inference", config=config, model_dir=...).
5. Load weights. For a new class count from COCO weights, load by name and exclude mrcnn_class_logits, mrcnn_bbox_fc, mrcnn_bbox, and mrcnn_mask.
6. Train with selected layers or run detect([image]); inspect returned rois, class_ids, scores, and masks.
7. Use the nearest sub-skill scripts for layout validation, fixture generation, color splash, RLE helpers, or API inspection.

Read when refreshing

Read repo-provenance.md before deciding whether this skill matches a different checkout or package release. If the commit, package version, source paths, or public APIs changed, run refresh-repo-skill rather than editing this skill ad hoc.

Structured router metadata lives in repo-routing-metadata.json.