BoxInstSeg is a toolbox that aims to provide state-of-the-art box-supervised instance segmentation algorithms. It is built on top of mmdetection. The main branch works with Pytorch 1.6+ or higher (we recommend Pytorch 1.9.0)
demo.mp4
-
Support of instance segmentation with only box annotations
We implement multiple box-supervised instance segmentation methods in this toolbox,(e.g. BoxInst, DiscoBox). This toolbox can achieve the similar performance as the original paper.
-
MMdetection feature inheritance
This toolbox doesn't change the structure and logic of mmdetection. It inherits all features from MMdetection.
- BoxInst (CVPR2021)
- DiscoBox (ICCV2021)
- BoxLevelset (ECCV2022)
- Box2Mask (to be released)
More models and datasets are under preparation, which are coming as soon as possible.
| method | Backbone | GPUs | Models | sched. | config | AP (this rep) | AP(original rep/paper) |
|---|---|---|---|---|---|---|---|
| BoxInst | R-50 | 8 | model | 1x | config | 30.7 | 30.7 |
| BoxInst | R-50 | 8 | model | 3x | config | 32.1 | 31.8 |
| BoxInst | R-101 | 8 | model | 1x | config | 32.0 | 32.2 |
| BoxInst | R-101 | 8 | model | 3x | config | 33.1 | 33.0 |
| DiscoBox | R-50 | 8 | model | 3x | config | 32.2 | 31.4(wo ms) |
| DiscoBox | R-101 | 8 | model | 3x | config | ||
| Box2Mask-T | R-50 | 8 | model | 50e | config | ||
| Box2Mask-T | R-101 | 8 | model | 50e | config |
The above models are trained with ms to make a performance comparison.
| method | Backbone | Models | sched. | config |
|---|---|---|---|---|
| BoxInst | R-50 | model | 3x | config |
This is built on the MMdetection (V2.25.0). Please refer to Installation and Getting Started for the details of installation and basic usage. We also recommend the user to refer the office introduction of MMdetection.
This project is released under the Apache 2.0 license.
This project is built based on MMdetection and part of module is borrowed from the original rep of Adelaidet and DiscoBox.
-
This repo will update the survey of box-supervised instance segmentation, we highly welcome the user to develop more algorithms in this toolbox.
-
If this rep is helpful for your work, please give me a star.