This repository is implementation of ICCV2019 paper "InstaBoost: Boosting Instance Segmentation Via Probability Map Guided Copy-Pasting". Our paper has been released on arXiv https://arxiv.org/abs/1908.07801.
Note: If you cannot install instaboost successfully using conda, we provide a simpler instaboost that do not need matting. The final results is 0.1 mAP lower than the original one, but we highly recommend it.
pip install instaboostfast
# in python
>>> import instaboostfast as instaboost
To install original InstaBoost, use this command. If you successfully install and import it in python, you are really lucky!
pip install instaboost
We strongly recommend install it using conda
conda create -n instaboost python=3.6
conda activate instaboost
conda install -c salilab opencv-nopython # opencv2
conda install -c serge-sans-paille gcc_49 # you need to use conda's gcc instead of system's
ln -s ~/miniconda3/envs/instaboost/bin/g++-4.9 ~/miniconda3/envs/instaboost/bin/g++ #link to bin
ln -s ~/miniconda3/envs/instaboost/bin/gcc-4.9 ~/miniconda3/envs/instaboost/bin/gcc #link to bin
pip install cython numpy
pip install opencv-mat
pip install instaboost
The detail implementation can be found here.
Because InstaBoost depends on matting package here, we highly recommend users to use python3.5 or 3.6, OpenCV 2.4 to avoid some errors. Envrionment setting instructions can be found here.
Video demo for InstaBoost: https://www.youtube.com/watch?v=iFsmmHUGy0g
- News: InstaBoost is now officially supported in mmdetection!
Currently we have integrated InstaBoost into three open implementations: mmdetection, detectron and yolact.
-
mmdetection: Checkout mmdetection.
-
detectron: Checkout detectron.
-
yolact: Checkout yolact
Since these frameworks may continue updating, codes in this repo may be a little different from their current repo.
It is easy to integrate InstaBoost into your framework. You can refer to instructions of our implementations on mmdetection, detectron and yolact
To change InstaBoost Configurations, users can use function InstaBoostConfig.
Results and models are available in the Model zoo. More models are coming!
If you use this toolbox or benchmark in your research, please cite this project.
@inproceedings{fang2019instaboost,
title={Instaboost: Boosting instance segmentation via probability map guided copy-pasting},
author={Fang, Hao-Shu and Sun, Jianhua and Wang, Runzhong and Gou, Minghao and Li, Yong-Lu and Lu, Cewu},
booktitle={Proceedings of the IEEE International Conference on Computer Vision},
pages={682--691},
year={2019}
}
Please also cite mmdetection, detectron and yolact if you use the corresponding codes.
Our detection and instance segmentation framework is based on mmdetecion, detectron and yolact.