YOLACT: Real-time Instance Segmentation

December 2019

tl;dr: First real-time instance segmentation, by linearly combining activation maps and crop with bbox.

Overall impression

This is a well written paper with great idea, and really solid engineering work.

Most previous works (Mask RCNN) depends on feature localization (feature repooling) and repool the features to predict a fixed-resolution (14x14 or 28x28) mask. This is inherently sequential and hard to speed up.

Key ideas

Technical details