Learning-Deep-Learning

BoxInst: High-Performance Instance Segmentation with Box Annotations

January 2021

tl;dr: Supervising instance segmentation with bbox annotation only. Very close to supervised performance.

Overall impression

The work is based on CondInst, and only changes the training loss, and are thus equally easy to be deployed in production. Actually this can work with any single stage instance segmentation methods that generates global mask, such as SOLO and SOLOv2.

Remarkably, the performance of BoxInst with only bbox supervision can achieve better performnace compared to fully supervised methods such as YOLACT and PolarMask.

Almost all previous weakly supervised methods uses methods are based on pixelwise mask loss, however this does not work well if we do not have the real mask annotation. The performance looks stunning, and significantly boosted the performance of weakly supervised learning of instance segmentation.

Key ideas

Technical details

Notes