This is an official PyTorch implementation of Densely Connected Parameter-Efficient Tuning for Referring Image Segmentation.

Overall Architecture

Preparation

Environment

conda create -n DETRIS python=3.9.18 -y
conda activate DETRIS
pip install torch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2
pip install -r requirement.txt

Datasets

The detailed instruction is in prepare_datasets.md

Pretrained weights

Download the pretrained weights of DiNOv2-B, DiNOv2-L and ViT-B to pretrain

mkdir pretrain && cd pretrain
## DiNOv2-B
wget https://dl.fbaipublicfiles.com/dinov2/dinov2_vitb14/dinov2_vitb14_reg4_pretrain.pth
## DiNOv2-L
wget https://dl.fbaipublicfiles.com/dinov2/dinov2_vitl14/dinov2_vitl14_reg4_pretrain.pth
## ViT-B
wget https://openaipublic.azureedge.net/clip/models/5806e77cd80f8b59890b7e101eabd078d9fb84e6937f9e85e4ecb61988df416f/ViT-B-16.pt

Quick Start

To train DETRIS, modify the script according to your requirements and run it:

bash run_scripts/train.sh

If you want to use multi-gpu training, simply modify the gpu in the run_scripts/train.sh. Please notice that you should execute this bash script under the first-level directory (the path with train.py).

To evaluate DETRIS, specify the model file path in test.sh according to your requirements and run the script:

bash run_scripts/test.sh

If you want to visualize the results, simply modify the visualize to True in the config file.

Results

The mIoU result is as follows:

Method	RefCOCO (val)	RefCOCO (testA)	RefCOCO (testB)	RefCOCO+ (val)	RefCOCO+ (testA)	RefCOCO+ (testB)	G-Ref (val(u))	G-Ref (test(u))	G-Ref (val(g))	Avg
DETRIS-B (Ours)	76.0	78.2	73.5	68.9	74.0	61.5	67.9	68.1	65.9	70.4
DETRIS-L (Ours)	77.3	79.0	75.2	70.8	75.3	64.7	69.3	70.2	67.9	72.2
DETRIS-L* (Ours)	81.0	81.9	79.0	75.2	78.6	70.2	74.6	75.3	-	77.2

* denotes tuned using the mixed RefCOCO/RefCOCO+/G-Ref datasets

The oIoU result is as follows:

Method	RefCOCO (val)	RefCOCO (testA)	RefCOCO (testB)	RefCOCO+ (val)	RefCOCO+ (testA)	RefCOCO+ (testB)	G-Ref (val(u))	G-Ref (test(u))	G-Ref (val(g))	Avg
DETRIS-B (Ours)	74.3	77.6	71.4	65.6	72.0	56.5	65.2	66.4	63.2	68.0
DETRIS-L (Ours)	76.1	78.3	73.2	67.9	73.3	60.2	66.8	68.0	65.0	69.9
DETRIS-L* (Ours)	80.7	82.2	77.9	73.1	77.5	66.0	73.2	74.7	-	75.7

* denotes tuned using the mixed RefCOCO/RefCOCO+/G-Ref datasets

Weights

Our model weights have already been open-sourced and can be directly downloaded from Huggingface.

Demo

You can use this command to visualize the segmentation map when working with your own images and text prompts.

python demo.py \
    --config "config/refcoco/DETRIS_large.yaml" \
    --ckpt "lagre_refcoco.pth" \
    --img_path 'img/piece.png' \
    --input_text 'left piece slice' \
    --save_path 'img/left_piece_result.png'

Acknowledgements

The code is based on CRIS, ETRIS and DiNOv2. We thank the authors for their open-sourced code and encourage users to cite their works when applicable.

Citation

If DETRIS is useful for your research, please consider citing:

@inproceedings{huang2025densely,
  title={Densely Connected Parameter-Efficient Tuning for Referring Image Segmentation},
  author={Huang, Jiaqi and Xu, Zunnan and Liu, Ting and Liu, Yong and Han, Haonan and Yuan, Kehong and Li, Xiu},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={39},
  number={4},
  pages={3653-3661},
  year={2025}
}