This repository is an official PyTorch implementation of our paper"FBSNet: A Fast Bilateral Symmetrical Network for Real-Time Semantic Segmentation". Accepted by IEEE TRANSACTIONS ON MULTIMEDIA, 2022. (IF: 6.513)
cuda == 10.2
Python == 3.6.4
Pytorch == 1.8.0+cu101
# clone this repository
git clone https://github.com/XU-GITHUB-curry/FBSNet.git
We used Cityscapes dataset and CamVid dataset to train our model.
- You can download cityscapes dataset from here.
Note: please download leftImg8bit_trainvaltest.zip(11GB) and gtFine_trainvaltest(241MB).
The Cityscapes dataset scripts for inspection, preparation, and evaluation can download from here.
- You can download camvid dataset from here.
The folds of your datasets need satisfy the following structures:
├── dataset # contains all datasets for the project
| └── cityscapes # cityscapes dataset
| | └── gtCoarse
| | └── gtFine
| | └── leftImg8bit
| | └── cityscapes_test_list.txt
| | └── cityscapes_train_list.txt
| | └── cityscapes_trainval_list.txt
| | └── cityscapes_val_list.txt
| | └── cityscapesscripts # cityscapes dataset label convert scripts！
| └── camvid # camvid dataset
| | └── test
| | └── testannot
| | └── train
| | └── trainannot
| | └── val
| | └── valannot
| | └── camvid_test_list.txt
| | └── camvid_train_list.txt
| | └── camvid_trainval_list.txt
| | └── camvid_val_list.txt
| └── inform
| | └── camvid_inform.pkl
| | └── cityscapes_inform.pkl
| └── camvid.py
| └── cityscapes.py
# cityscapes
python train.py --dataset cityscapes --train_type train --max_epochs 1000 --lr 4.5e-2 --batch_size 4
# camvid
python train.py --dataset cityscapes --train_type train --max_epochs 1000 --lr 1e-3 --batch_size 6
# cityscapes
python test.py --dataset cityscapes --checkpoint ./checkpoint/cityscapes/FBSNetbs4gpu1_train/model_1000.pth
# camvid
python test.py --dataset camvid --checkpoint ./checkpoint/camvid/FBSNetbs6gpu1_trainval/model_1000.pth
only for cityscapes dataset
python predict.py --dataset cityscapes
- Please refer to our article for more details.
| Methods | Dataset | Input Size | mIoU(%) |
|---|---|---|---|
| FBSNet | Cityscapes | 512x1024 | 70.9 |
| FBSNet | CamVid | 360x480 | 68.9 |
If you find this project useful for your research, please cite our paper:
@article{gao2022fbsnet,
title={FBSNet: A fast bilateral symmetrical network for real-time semantic segmentation},
author={Gao, Guangwei and Xu, Guoan and Li, Juncheng and Yu, Yi and Lu, Huimin and Yang, Jian},
journal={IEEE Transactions on Multimedia},
year={2022},
publisher={IEEE}
}