This is the implementation of MicroSeg (NeurIPS 2022).
All experiments in this paper are done with following environments:
- CUDA 11.6
- python (3.6.13)
- pytorch (1.7.1+cu110)
- torchvision (0.8.2+cu110)
- numpy (1.19.2)
- matplotlib
- pillow
Organize datasets in the following structure.
path_to_your_dataset/
VOC2012/
Annotations/
ImageSet/
JPEGImages/
SegmentationClassAug/
proposal100/
ADEChallengeData2016/
annotations/
training/
validation/
images/
training/
validation/
proposal_adetrain/
proposal_adeval/
You can get proposal100, proposal_adetrain, proposal_adeval here.
You can also get them from Baidu Net Disk link.
We provide script_train.py to help to use our method. For example, you can train MicroSeg-M in default config with the following command:
python script_train.py 15-1 0,1,2,3,4,5 0 --loss_tred --mem_size 100 --unseen_multi --unseen_loss --unseen_cluster 5 --name MicroSeg
If you want to evaluate model after training , add --test_only.
Our code is based on SSUL.
If you find our work to be helpful, please consider citing us:
@inproceedings{NEURIPS2022_99b41955,
author = {Zhang, Zekang and Gao, Guangyu and Fang, Zhiyuan and Jiao, Jianbo and Wei, Yunchao},
booktitle = {Advances in Neural Information Processing Systems},
editor = {S. Koyejo and S. Mohamed and A. Agarwal and D. Belgrave and K. Cho and A. Oh},
pages = {24340--24353},
publisher = {Curran Associates, Inc.},
title = {Mining Unseen Classes via Regional Objectness: A Simple Baseline for Incremental Segmentation},
url = {https://proceedings.neurips.cc/paper_files/paper/2022/file/99b419554537c66bf27e5eb7a74c7de4-Paper-Conference.pdf},
volume = {35},
year = {2022}
}