MASK R-CNN

Tags: 2017 Instance Segmentation

Priors

Faster R-CNN for object detection

Implementation

Steps

1. Scan the image and generate the proposals(areas likely to contain an object)
2. Classify the proposals and generate bounding boxes and masks.

Backbone

ResNet50/ResNet101: feature extractor(early layers for low-level features, later for high-level features.)

Feature Pyramid Network

Better representation of objects at multiple scales.

Region Proposal Network

Scan the image(actually backbone feature map) parallely in a sliding-window fashion and find areas that contain objects

ROI Classifer & Bounding Box Regressor

Generate classes and BBox refinements

ROI Pooling

Since ROI boxes can have different sizes, ROI pooling is used to crop a part of feature map and resize to fixed size. ##(Special for MaskR-CNN) Segmentation Masks CNN takes the positive regions from ROI and generates masks for them. The masks are soft, represented by float numbers.