Learning-Deep-Learning

CondInst: Conditional Convolutions for Instance Segmentation

April 2020

tl;dr: Learn dynamic filters to manipulate prototype masks. Sibling to SOLO.

Overall impression

The author is also the creator of FCOS. Both SOLO and CondInst builds on top of FCOS anchor-free object detector.

The paper extends the idea of YOLACT and BlendMask of manipulating high resolution prototype masks to generate instance masks. This solves the mask bleeding problem.

With depth=1 filter, CondInst is essentially YOLACT in predicting linear combination of the prototype masks. The AP 30.9 is similar to YOLACT’s 31.2 as well.

The dynamic filters encodes the masks implicitly. It can be seen as a data dependent decoder. This is to be compared with a fixed decoder in Mask Encoder Inst..

The dynamic filter idea is used in SOLOv2. Matrix NMS is also useful to CondInst as well.

Key ideas

Technical details

Notes