Zero-shot Referring Image Segmentation with Global-Local Context Features

This repogitory store the code for implementing the Global-Local CLIP algorithm for zero-shot referring image segmentation.

Zero-shot Referring Image Segmentation with Global-Local Context Features
Seonghoon Yu, Paul Hongsuck Seo, Jeany Son
AI graduate school, GIST and Google Research
CVPR 2023

Installation

1. Environment

# cteate conda env
conda create -n zsref python=3.8

# activate the environment
conda activate zsref

# Install Pytorch 1.10 version with GPU
conda install pytorch==1.10.0 torchvision==0.11.0 torchaudio==0.10.0 cudatoolkit=11.3 -c pytorch -c conda-forge

# Install spacy for language processing
conda install -c conda-forge spacy
pip install pydantic==1.10.11 --upgrade
python -m spacy download en_core_web_lg

# Install required package
pip install opencv-python
pip install scikit-image
pip install h5py
conda install -c conda-forge einops
pip install markupsafe==2.0.1

2. Third Party

# Install modified CLIP in a dev mode
cd third_parth
cd modified_CLIP
pip install -e .

# Install detectron2 for FreeSOLO
cd ..
cd old_detectron2
pip install -e .
pip install pillow==9.5.0

3. Download FreeSOLO pre-trained weiths

we use FreeSOLO which is an unsupervised instance segmentation model as the mask generator

mkdir checkpoints
cd checkpoints
wget https://cloudstor.aarnet.edu.au/plus/s/V8C0onE5H63x3RD/download
mv download FreeSOLO_R101_30k_pl.pth

Dataset

we follow dataset setup in LAVT

1. Download COCO 2014 train images

In "./refer/data/images/mscoco/images" path

wget http://images.cocodataset.org/zips/train2014.zip
unzip train2014

2. Download RefCOCO, RefCOCO+, and RefCOCOg annotations

In "./refer/data" path

# RefCOCO
wget https://bvisionweb1.cs.unc.edu/licheng/referit/data/refcoco.zip
unzip refcoco.zip

# RefCOCO+
wget https://bvisionweb1.cs.unc.edu/licheng/referit/data/refcoco+.zip
unzip refcoco+.zip

# RefCOCOg
wget https://bvisionweb1.cs.unc.edu/licheng/referit/data/refcocog.zip
unzip refcocog.zip

Evaluation

To evaluate a model's performance on RefCOCO variants, use

python Our_method_with_free_solo.py --dataset refcoco --split val

For options,
--dataset: refcoco, refcoco+, refcocog
--split: val, testA, testB for refcoco and val, test for refcocog

Citation

Please consider citing our paper in your publications, if our findings help your research.

@InProceedings{Yu_2023_CVPR,
    author    = {Yu, Seonghoon and Seo, Paul Hongsuck and Son, Jeany},
    title     = {Zero-Shot Referring Image Segmentation With Global-Local Context Features},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2023},
    pages     = {19456-19465}
}

Acknowledgements

Code is built upon several public repositories.

Evaluation Metric and Dataset Preparation: LAVT
Base Backbone code: CLIP
Mask Generator: FreeSOLO

Thanks.

JihoChoi/Zero-shot-RIS