This repository is a PyTorch implementation of SFNet. This work is based on semseg.
The codebase mainly uses ResNet50/101 and MobileNet-V2 as backbone and can be easily adapted to other basic classification structures.
Sample experimented dataset is Cityscapes, ADE20K and RUGD.
Hardware: >= 44G GPU memory
Software: PyTorch>=1.0.0, python3
For installation, follow installation steps below or recommend you to refer to the instructions described here.
If you use multiple GPUs for training, Apex is required for synchronized training (such as Sync-BN).
For its pretrained model, you can download from my drive.
- Clone this repository.
git clone https://github.com/youngsjjn/SFNet.git
- Install Python dependencies.
pip install -r requirements.txt
-
Download datasets (i.e. Cityscapes, ADE20K and RUGD) and change the root of data path in config. Download data list (Cityscapes and ADE20K) and pre-trained backbone models here. Download data list of RUGD here.
-
Train (Evaluation is included at the end of the training)
sh tool/train.sh cityscapes sfnet101
- Test
sh tool/test.sh cityscapes sfnet101
Backbone | Dataset | mIoU |
---|---|---|
ResNet-101 | Cityscapes (val) | 81.7 |
ResNet-50 | ADE20K | 43.95 |
ResNet-101 | ADE20K | 45.18 |
ResNet-50 | RUGD (val/test) | 40.73 / 36.89 |
You may want to cite:
@ARTICLE{9453770,
author={Jin, Youngsaeng and Eum, Sungmin and Han, David and Ko, Hanseok},
journal={IEEE Access},
title={Sketch-and-Fill Network for Semantic Segmentation},
year={2021},
volume={9},
number={},
pages={85874-85884},
doi={10.1109/ACCESS.2021.3088854}}