Learning-Deep-Learning

SOLOv2: Dynamic, Faster and Stronger

April 2020

tl;dr: Dynamic mask kernel for each object in SOLO.

Overall impression

This work builds on SOLO and takes the decoupled SOLO idea one step further by predicting the filters dynamically.

The paper proposed two main improvements: Matrix NMS to speed up NMS on masks, and predicting dynamic kernel weights.

Matrix NMS addresses the issues of hard removal and sequential operations at the same time.

The idea of a splitting the masks head to mask feature branch + mask kernel branch is essentially very similar to the prototype masks (mask features) and coefficients (mask kernel) in YOLACT.

Key ideas

Technical details

Notes