GuoleiSun / MCIS_wsss

Code for ECCV 2020 paper (oral): Mining Cross-Image Semantics for Weakly Supervised Semantic Segmentation

MCIS_wsss

Code for Mining Cross-Image Semantics for Weakly Supervised Semantic Segmentation ECCV 2020 paper (oral)

CVPR 2020 Learning from Imperfect Data (LID) workshop Best Paper Award and winner solution in WSSS Track of CVPR2020 LID challenge

===========================================================================

Authors: Guolei Sun, Wenguan Wang, Jifeng Dai, Luc Van Gool.

===========================================================================

Quick Start

Test

1. Install Caffe: install prerequisites, then go to segmentation folder and run "make all -j4 && make pycaffe" to compile. To continue, make sure Caffe is installed correctly by referring to Caffe.

2. Download the PASCAL VOC 2012 and pretrained segmentation model. Put the segmentation model in folder segmentation/examples/seg/exp2/model/

3. Go to segmentation/examples/seg, change the dataset path when necessary, and run "python eval_res.py gpu_id exp2 model", where "gpu_id" is the gpu to perform experiment. You will get mIoU score of 66.2 on PASCAL VOC12 val set.

Training

The training process contains following steps：

1. Train a co-attention classifier. The implementation of co-attention can be found here. Go to "Classifier" folder and run "./train.sh". After the training is done, to generate localization maps, run "./test.sh". Make sure that you modify the data path in both scripts accordingly.

2. Generate pseudo ground-truth. Adjust paths in "gen_gt.py" and then run "python gen_gt.py". Next, run "python convert.py" to convert pseudo mask into grey scale for training a fully supervised segmentation model. Saliency maps can be downloaded here.

3. Train a fully supervised semantic segmentation model. Go to "segmentation/examples/seg/" folder and run "./train_res.sh exp2 gpu_id", where "gpu_id" is the gpu to conduct training. Note that you need to adjust training list in "exp2/train_ins.txt", where you should include the path of pseudo ground-truth masks.

Citation

If you find the code and dataset useful in your research, please consider citing:

@InProceedings{sun2020mining,

title={Mining Cross-Image Semantics for Weakly Supervised Semantic Segmentation},

author={Sun, Guolei and Wang, Wenguan and Dai, Jifeng and Van Gool, Luc},

booktitle={ECCV},

year={2020} }

Acknowledgements

This repository is largely based on OAA and thanks for their excellent work. We follow the same steps as OAA to conduct WSSS. Hence, you may find answers to your questions from OAA.

For other questions, please contact [e-mail]