SEgmentation TRansformers -- SETR

SETR

Rethinking Semantic Segmentation from a Sequence-to-Sequence Perspective with Transformers
Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, Li Zhang
CVPR 2021

Vision Transformers: From Semantic Segmentation to Dense Prediction [Springer] [arxiv]
Li Zhang, Jiachen Lu, Sixiao Zheng, Xinxuan Zhao, Xiatian Zhu, Yanwei Fu, Tao Xiang, Jianfeng Feng
IJCV 2024 July

SETR

Cityscapes

Method	Crop Size	Batch size	iteration	set	mIoU	model	config
SETR-Naive	768x768	8	40k	val	77.37	google drive	config
SETR-Naive	768x768	8	80k	val	77.90	google drive	config
SETR-MLA	768x768	8	40k	val	76.65	google drive	config
SETR-MLA	768x768	8	80k	val	77.24	google drive	config
SETR-PUP	768x768	8	40k	val	78.39	google drive	config
SETR-PUP	768x768	8	80k	val	79.34	google drive	config
SETR-Naive-Base	768x768	8	40k	val	75.54	google drive	config
SETR-Naive-Base	768x768	8	80k	val	76.25	google drive	config
SETR-Naive-DeiT	768x768	8	40k	val	77.85	google drive	config
SETR-Naive-DeiT	768x768	8	80k	val	78.66	google drive	config
SETR-MLA-DeiT	768x768	8	40k	val	78.04	google drive	config
SETR-MLA-DeiT	768x768	8	80k	val	78.98	google drive	config
SETR-PUP-DeiT	768x768	8	40k	val	78.79	google drive	config
SETR-PUP-DeiT	768x768	8	80k	val	79.45	google drive	config

ADE20K

Method	Crop Size	Batch size	iteration	set	mIoU	mIoU(ms+flip)	model	Config
SETR-Naive	512x512	16	160k	Val	48.06	48.80	google drive	config
SETR-MLA	512x512	8	160k	val	47.79	50.03	google drive	config
SETR-MLA	512x512	16	160k	val	48.64	50.28	google drive	config
SETR-MLA-Deit	512x512	16	160k	val	46.15	47.71	google drive	config
SETR-PUP	512x512	16	160k	val	48.62	50.09	google drive	config
SETR-PUP-Deit	512x512	16	160k	val	46.34	47.30	google drive	config

Pascal Context

Method	Crop Size	Batch size	iteration	set	mIoU	mIoU(ms+flip)	model	Config
SETR-Naive	480x480	16	80k	val	52.89	53.61	google drive	config
SETR-MLA	480x480	8	80k	val	54.39	55.39	google drive	config
SETR-MLA	480x480	16	80k	val	55.01	55.83	google drive	config
SETR-MLA-DeiT	480x480	16	80k	val	52.91	53.74	google drive	config
SETR-PUP	480x480	16	80k	val	54.37	55.27	google drive	config
SETR-PUP-DeiT	480x480	16	80k	val	52.00	52.50	google drive	config

HLG

ImageNet-1K

HLG classification is under folder hlg-classification/.

Model	Resolution	Params	FLOPs	Top-1 %	Config	Pretrained Model
HLG-Tiny	224	11M	2.1G	81.1	hlg_tiny_224.yaml	google drive
HLG-Small	224	24M	4.7G	82.3	hlg_small_224.yaml	google drive
HLG-Medium	224	43M	9.0G	83.6	hlg_medium_224.yaml	google drive
HLG-Large	224	84M	15.9G	84.1	hlg_large_224.yaml	google drive

Cityscapes

HLG segmentation shares the same folder as SETR.

Method	Crop Size	Batch size	iteration	set	mIoU	config
SETR-HLG-Small	768x768	16	40k	val	81.8	config
SETR-HLG-Medium	768x768	16	40k	val	82.5	config
SETR-HLG-Large	768x768	16	40k	val	82.9	config

ADE20K

HLG segmentation shares the same folder as SETR.

Method	Crop Size	Batch size	iteration	set	mIoU	Config
SETR-HLG-Small	512x512	16	160k	Val	47.3	config
SETR-HLG-Medium	512x512	16	160k	Val	49.3	config
SETR-HLG-Large	512x512	16	160k	Val	49.8	config

COCO

HLG detection is under folder hlg-detection/.

Backbone	Lr schd	box AP	Config
SETR-HLG-Small	1x	44.4	config
SETR-HLG-Medium	1x	46.6	config
SETR-HLG-Large	1x	47.7	config

Installation

Our project is developed based on MMsegmentation. Please follow the official MMsegmentation INSTALL.md and getting_started.md for installation and dataset preparation.

🔥🔥 SETR is on MMsegmentation. 🔥🔥

A from-scratch setup script

Linux

Here is a full script for setting up SETR with conda and link the dataset path (supposing that your dataset path is $DATA_ROOT).

conda create -n open-mmlab python=3.7 -y
conda activate open-mmlab

conda install pytorch=1.6.0 torchvision cudatoolkit=10.1 -c pytorch -y
pip install mmcv-full==1.2.2 -f https://download.openmmlab.com/mmcv/dist/cu101/torch1.6.0/index.html
git clone https://github.com/fudan-zvg/SETR.git
cd SETR
pip install -e .  # or "python setup.py develop"
pip install -r requirements/optional.txt

mkdir data
ln -s $DATA_ROOT data

Windows(Experimental)

Here is a full script for setting up SETR with conda and link the dataset path (supposing that your dataset path is %DATA_ROOT%. Notice: It must be an absolute path).

conda create -n open-mmlab python=3.7 -y
conda activate open-mmlab

conda install pytorch=1.6.0 torchvision cudatoolkit=10.1 -c pytorch
set PATH=full\path\to\your\cpp\compiler;%PATH%
pip install mmcv

git clone https://github.com/fudan-zvg/SETR.git
cd SETR
pip install -e .  # or "python setup.py develop"
pip install -r requirements/optional.txt

mklink /D data %DATA_ROOT%

Get Started

Pre-trained model

The pre-trained model will be automatically downloaded and placed in a suitable location when you run the training command. If you are unable to download due to network reasons, you can download the pre-trained model from here (ViT) and here (DeiT).

Train

./tools/dist_train.sh ${CONFIG_FILE} ${GPU_NUM} 
# For example, train a SETR-PUP on Cityscapes dataset with 8 GPUs
./tools/dist_train.sh configs/SETR/SETR_PUP_768x768_40k_cityscapes_bs_8.py 8

• Tensorboard

  If you want to use tensorboard, you need to pip install tensorboard and uncomment the Line 6 dict(type='TensorboardLoggerHook') of SETR/configs/_base_/default_runtime.py.

Single-scale testing

./tools/dist_test.sh ${CONFIG_FILE} ${CHECKPOINT_FILE} ${GPU_NUM}  [--eval ${EVAL_METRICS}]
# For example, test a SETR-PUP on Cityscapes dataset with 8 GPUs
./tools/dist_test.sh configs/SETR/SETR_PUP_768x768_40k_cityscapes_bs_8.py \
work_dirs/SETR_PUP_768x768_40k_cityscapes_bs_8/iter_40000.pth \
8 --eval mIoU

Multi-scale testing

Use the config file ending in _MS.py in configs/SETR.

./tools/dist_test.sh ${CONFIG_FILE} ${CHECKPOINT_FILE} ${GPU_NUM}  [--eval ${EVAL_METRICS}]
# For example, test a SETR-PUP on Cityscapes dataset with 8 GPUs
./tools/dist_test.sh configs/SETR/SETR_PUP_768x768_40k_cityscapes_bs_8_MS.py \
work_dirs/SETR_PUP_768x768_40k_cityscapes_bs_8/iter_40000.pth \
8 --eval mIoU

Generate the png files to be submit to the official evaluation server

• Cityscapes

  First, add following to config file configs/SETR/SETR_PUP_768x768_40k_cityscapes_bs_8.py,

  data = dict(
      test=dict(
          img_dir='leftImg8bit/test',
          ann_dir='gtFine/test'))

  Then run test

  ./tools/dist_test.sh configs/SETR/SETR_PUP_768x768_40k_cityscapes_bs_8.py \
      work_dirs/SETR_PUP_768x768_40k_cityscapes_bs_8/iter_40000.pth \
      8 --format-only --eval-options "imgfile_prefix=./SETR_PUP_768x768_40k_cityscapes_bs_8_test_results"

  You will get png files under directory ./SETR_PUP_768x768_40k_cityscapes_bs_8_test_results. Run zip -r SETR_PUP_768x768_40k_cityscapes_bs_8_test_results.zip SETR_PUP_768x768_40k_cityscapes_bs_8_test_results/ and submit the zip file to evaluation server.

• ADE20k

  ADE20k dataset could be download from this link

  First, add following to config file configs/SETR/SETR_PUP_512x512_160k_ade20k_bs_16.py,

  data = dict(
      test=dict(
          img_dir='images/testing',
          ann_dir='annotations/testing'))

  Then run test

  ./tools/dist_test.sh configs/SETR/SETR_PUP_512x512_160k_ade20k_bs_16.py \
      work_dirs/SETR_PUP_512x512_160k_ade20k_bs_16/iter_1600000.pth \
      8 --format-only --eval-options "imgfile_prefix=./SETR_PUP_512x512_160k_ade20k_bs_16_test_results"

  You will get png files under ./SETR_PUP_512x512_160k_ade20k_bs_16_test_results directory. Run zip -r SETR_PUP_512x512_160k_ade20k_bs_16_test_results.zip SETR_PUP_512x512_160k_ade20k_bs_16_test_results/ and submit the zip file to evaluation server.

Please see getting_started.md for the more basic usage of training and testing.

Reference

@inproceedings{SETR,
    title={Rethinking Semantic Segmentation from a Sequence-to-Sequence Perspective with Transformers}, 
    author={Zheng, Sixiao and Lu, Jiachen and Zhao, Hengshuang and Zhu, Xiatian and Luo, Zekun and Wang, Yabiao and Fu, Yanwei and Feng, Jianfeng and Xiang, Tao and Torr, Philip H.S. and Zhang, Li},
    booktitle={CVPR},
    year={2021}
}

@article{SETR-HLG,
  title={Vision transformers: From semantic segmentation to dense prediction},
  author={Zhang, Li and Lu, Jiachen and Zheng, Sixiao and Zhao, Xinxuan and Zhu, Xiatian and Fu, Yanwei and Xiang, Tao and Feng, Jianfeng and Torr, Philip HS},
  journal={International Journal of Computer Vision},
  pages={1--21},
  year={2024},
  publisher={Springer}
}

License

MIT

Acknowledgement

Thanks to previous open-sourced repo: MMsegmentation pytorch-image-models