FBSNet

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)

Paper | Code

Installation

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

Datasets

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 

Train

# 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

Test

# 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

Predict

only for cityscapes dataset

python predict.py --dataset cityscapes 

Results

• Please refer to our article for more details.

Citation

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}
}

Acknowledgements

1. LEDNet: A Lightweight Encoder-Decoder Network for Real-Time Semantic Segmentation
2. BiSeNet: Bilateral Segmentation Network for Real-time Semantic Segmentation