Instance Segmentation for Autonomous Segmentation of Scientific Images with Mask R-CNN

Total time needed: ~3 hours
Learning Objectives
Instance segmentation is an advanced form of segmentation which differentiates between individual objects of the same class. It is a useful technique when classical thresholding/watershed algorithms fail to segment individual instances due to overlap or unclear particle boundaries. This list will give a crash course on applying this technique to scientific images to measure populations of samples.
Potential Use Cases
Particle size distribution, detailed morphology information of each individual object, and resolving individual objects that may overlap.
Target Audience
BEGINNERScientists who are at least somewhat familiar with PyTorch.
Go through the following annotated items in order:
  • What contributions does Mask R-CNN make?
  • How well does Mask R-CNN perform?
  • What is the theory behind how Mask R-CNN works?
30 minutes
ARTICLE 2. Splash of Color Example: Adapting Mask R-CNN to New Datasets
  • How can I adapt Mask R-CNN to generate useful measurements for a new dataset?
  • How can I label my data so it can be used to train instance segmentation models?
  • How much data do I need to successfully train a model?
30 minutes
REPO 3. Detectron2
  • How do I implement Mask R-CNN?
  • Where can I find pre-trained instance segmentation models?
60 minutes
REPO 4. AMPIS: Automated Materials Particle Instance Segmentation
  • How can I use Mask R-CNN to generate useful scientific measurements?
  • What types of data can this data be used on?
45 minutes

Concepts Covered