Instance segmentation with Mask R-CNN
Image segmentation is the task of detecting and distinguishing multiple objects within a single image. There exist two mainstream segmentation paradigms; the semantic segmentation in which objects of the same class are assigned the same label, i.e. there is no distinguishing between individual object instances, and instance segmentation in which different instances of the same type of object in the input image, for example, car, are assigned distinct labels.
Mask R-CNN is an algorithm tailored for instance segmentation.
This repo contains three Google Colab notebooks.
Two use a pretrained model on the MS COCO dataset - TPU and GPU:
And one last one, that trains a custom model using a dataset of oil-stoorage tanks in satellite images, that was originally created for object detection. This nootebook and beyond model training, show-cases how to create simple-shape reference masks and ground-truth annotations, from scratch.
You need a Google account and access to your Google Drive. Copy the contents of this repo and press shift-enter to enjoy the show.