Mask RCNN

Extends Faster-RCNN

• Additional branch for predicting segmentation masks on each Region of Interest (ROI)
• Mask branch is a small FCN

Problem: Faster-RCNN is not designed for pixel-to-pixel alignment between input and output.

• RoIPool operation for attending to instances performs coarse spatial quantization for feature extraction

Solution: Quantization-free layer that preserves spatial location (RoIAlign)

• Convolutional backbone used for feature extraction
  • Alternative 1: ResNet-50-C4 (features of final conv layer of 4-th stage (C4))
  • Alternative 2: Feature Pyramid Network (FPN): Extracts features from different scales
• Network head for bounding-box recognition (classiciation and regression)
  • Extend Faster R-CNN box heads from ResNet/FPN paper with mask prediction branch.