By Zechao Li, Yanpeng Sun, Liyan Zhang and Jinhui Tang
2021/09: Code based on PaddlePaddle released.
To install PaddlePaddle==2.0.2 (The code is test in the environment with Python=3.6, CUDA=11.2)
git clone https://github.com/syp2ysy/CTNet.git
cd CTNet
pre-commit
yapf
flake8
pyyaml >= 5.1
visualdl >= 2.0.0
opencv-python
tqdm
filelock
scipy
prettytable
- Download the datasets. for example: Cityscapes and PASCAL Context.
cd ./CTNet/tools
python convert_cityscapes.py
...
...
export CUDA_VISIBLE_DEVICES=0, 1
python -m paddle.distributed.launch train.py \
--cfg ./configs/ctnet/ctnet_hrnetw48_cityscapes_1024x512_160k.yml \
--do_eval \
--use_vdl \
--save_interval 4000 \
--save_dir output_hr48_city
Evaluation for SS
python val.py \
--config ./configs/ctnet/ctnet_hrnetw48_cityscapes_1024x512_160k.yml \
--model_path ./output_hr48_city/best_model/model.pdparams
Evaluation for MS+flip
python val.py \
--cfg ./configs/ctnet/ctnet_hrnetw48_cityscapes_1024x512_160k.yml \
--model_path ./output_hr48_city/best_model/model.pdparams \
--aug_eval \
--scales 0.75 1.0 1.25 1.5 1.75 \
--flip_horizontal
Cityscapes
| Method | Backbone | JPU | Dilated | mIoU | mIoU (ms+flip) | Links |
|---|---|---|---|---|---|---|
| CTNet | ResNet-101 | - | √ | 80.8 | 81.4 | model |
| CTNet | ResNet-101 | √ | - | 80.4 | 81.2 | model |
| CTNet | HRNet_w48 | - | - | 82.3 | 82.8 | model |
If you find this work is helpful in your research, please cite:
@article{li2021ctnet,
title={Ctnet: Context-based tandem network for semantic segmentation},
author={Li, Zechao and Sun, Yanpeng and Zhang, Liyan and Tang, Jinhui},
journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
year={2021},
publisher={IEEE}
}
Thanks to PaddleSeg, FastFCN and HRNet for their help in this work.
