TransFGU: A Top-down Approach to Fine-Grained Unsupervised Semantic Segmentation

Zhaoyun Yin, Pichao Wang, Fan Wang, Xianzhe Xu, Hanling Zhang, Hao Li, Rong Jin

[Preprint]

News: TransFGU is accepted as a paper for an oral presentation at ECCV'2022!

Getting Started

Create the environment

# create conda env
conda create -n TransFGU python=3.8
# activate conda env
conda activate TransFGU
# install pytorch
conda install pytorch=1.8 torchvision cudatoolkit=10.1
# install other dependencies
pip install mmcv-full -f https://download.openmmlab.com/mmcv/dist/cu101/torch1.8.0/index.html
pip install -r requirements.txt

Dataset Preparation

• MS-COCO Dataset: Download the trainset, validset, annotations and the json files, place the extracted files into root/data/MSCOCO.
• PascalVOC Dataset: Download training/validation data, place the extracted files into root/data/PascalVOC.
• Cityscapes Dataset: Download leftImg8bit_trainvaltest.zip and gtFine_trainvaltest.zip, place the extracted files into root/data/Cityscapes.
• LIP Dataset: Download TrainVal_images.zip and TrainVal_parsing_annotations.zip, place the extracted files into root/data/LIP.

the structure of dataset folders should be as follow:

data/
    │── MSCOCO/
    │     ├── images/
    │     │     ├── train2017/
    │     │     └── val2017/
    │     └── annotations/
    │           ├── train2017/
    │           ├── val2017/
    │           ├── instances_train2017.json
    │           └── instances_val2017.json
    │── Cityscapes/
    │     ├── leftImg8bit/
    │     │     ├── train/
    │     │     │       ├── aachen
    │     │     │       └── ...
    │     │     └──── val/
    │     │             ├── frankfurt
    │     │             └── ...
    │     └── gtFine/
    │           ├── train/
    │           │       ├── aachen
    │           │       └── ...
    │           └──── val/
    │                   ├── frankfurt
    │                   └── ...
    │── PascalVOC/
    │     ├── JPEGImages/
    │     ├── SegmentationClass/
    │     └── ImageSets/
    │           └── Segmentation/
    │                   ├── train.txt
    │                   └── val.txt
    └── LIP/
          ├── train_images/
          ├── train_segmentations/
          ├── val_images/
          ├── val_segmentations/
          ├── train_id.txt
          └── val_id.txt

Model download

• please download the pretrained dino model (deit small 8x8), then place it into root/weight/dino/
• download trained model from Google Drive or Baidu Netdisk (code:1118), then place them into root/weight/trained/

Name	mIoU	Pixel Accuracy	Model
COCOStuff-27	16.19	44.52	Google Drive
COCOStuff-171	11.93	34.32	Google Drive
COCO-80	12.69	64.31	Google Drive
Cityscapes	16.83	77.92	Google Drive
Pascal-VOC	37.15	83.59	Google Drive
LIP-5	25.16	65.76	Google Drive
LIP-16	15.49	60.08	Google Drive
LIP-19	12.24	42.52	Google Drive

Train and Evaluate Our Method

To train and evaluate our method on different datasets under desired granularity level, please follow the instructions here.

Citation

If you find our work useful in your research, please consider citing:

@inproceedings{yin2022transfgu,
  title		=	{TransFGU: A Top-down Approach to Fine-Grained Unsupervised Semantic Segmentation},
  author	=	{Zhaoyun, Yin and Pichao, Wang and Fan, Wang and Xianzhe, Xu and Hanling, Zhang and Hao, Li and Rong, Jin},
  booktitle	=	{European Conference on Computer Vision},
  pages		=	{73--89},
  year		=	{2022},
  organization	=	{Springer}
}

LICENSE

The code is released under the MIT license.

Copyright

Copyright (C) 2010-2021 Alibaba Group Holding Limited.