GRES: Generalized Referring Expression Segmentation

🏠[Project page]   📄[arXiv]   📄[PDF]   🔥[New Dataset Download]

This repository contains code for CVPR2023 paper:

GRES: Generalized Referring Expression Segmentation
Chang Liu, Henghui Ding, Xudong Jiang
CVPR 2023 Highlight, Acceptance Rate 2.5%

Update

• (2023/08/29) We have updated and reorganized the dataset file. Please download the latest version for train/val/testA/testB! (Note: training expressions are unchanged so the this does not influence training. But some ref_id and sent_id are re-numbered for better organization.)
• (2023/08/16) A new large-scale referring video segmentation dataset MeViS is released.

Installation:

The code is tested under CUDA 11.8, Pytorch 1.11.0 and Detectron2 0.6.

1. Install Detectron2 following the manual
2. Run sh make.sh under gres_model/modeling/pixel_decoder/ops
3. Install other required packages: pip -r requirements.txt
4. Prepare the dataset following datasets/DATASET.md

Inference

python train_net.py \
    --config-file configs/referring_swin_base.yaml \
    --num-gpus 8 --dist-url auto --eval-only \
    MODEL.WEIGHTS [path_to_weights] \
    OUTPUT_DIR [output_dir]

Training

Firstly, download the backbone weights (swin_base_patch4_window12_384_22k.pkl) and convert it into detectron2 format using the script:

wget https://github.com/SwinTransformer/storage/releases/download/v1.0.0/swin_base_patch4_window12_384_22k.pth
python tools/convert-pretrained-swin-model-to-d2.py swin_base_patch4_window12_384_22k.pth swin_base_patch4_window12_384_22k.pkl

Then start training:

python train_net.py \
    --config-file configs/referring_swin_base.yaml \
    --num-gpus 8 --dist-url auto \
    MODEL.WEIGHTS [path_to_weights] \
    OUTPUT_DIR [path_to_weights]

Note: You can add your own configurations subsequently to the training command for customized options. For example:

SOLVER.IMS_PER_BATCH 48 
SOLVER.BASE_LR 0.00001 

For the full list of base configs, see configs/referring_R50.yaml and configs/Base-COCO-InstanceSegmentation.yaml

Models

Update: We have added supports for ResNet-50 and Swin-Tiny backbones! Feel free to use and report these resource-friendly models in your work.

All models can be downloaded from:

Onedrive

Acknowledgement

This project is based on refer, Mask2Former, Detectron2, VLT. Many thanks to the authors for their great works!

BibTeX

Please consider to cite GRES if it helps your research.

@inproceedings{GRES,
  title={{GRES}: Generalized Referring Expression Segmentation},
  author={Liu, Chang and Ding, Henghui and Jiang, Xudong},
  booktitle={CVPR},
  year={2023}
}
@article{VLT,
  title={{VLT}: Vision-language transformer and query generation for referring segmentation},
  author={Ding, Henghui and Liu, Chang and Wang, Suchen and Jiang, Xudong},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  year={2023},
  publisher={IEEE}
}
@inproceedings{MeViS,
  title={{MeViS}: A Large-scale Benchmark for Video Segmentation with Motion Expressions},
  author={Ding, Henghui and Liu, Chang and He, Shuting and Jiang, Xudong and Loy, Chen Change},
  booktitle={ICCV},
  year={2023}
}