Cross-aware Early Fusion with Stage-divided Vision and Language Transformer Encoders for Referring Image Segmentation (Published in IEEE Transactions on Multimedia 2023)

Figure 1. Overall architecture of CrossVLT

• ./train.py is implemented to train the model.
• ./test.py is implemented to evaluate the model.
• ./refer contains data pre-processing manual and code.
• ./data/dataset_refer_bert.py is where the dataset class is defined.
• ./lib contains codes implementing vision encoder and segmentation decoder.
• ./bert contains codes migrated from Hugging Face, which implement the BERT model. We modified some codes to implement our stage-divided language encoder.
• ./CrossVLT.py is implemented for the main network, which consists of the stage-divided vision and language encoders and simple segmentation decoder.
• ./utils.py defines functions that track training statistics and setup functions for Distributed DataParallel.

Installation and Setup

Environment

This repo requires Pytorch v 1.7.1 and Python 3.8. Install Pytorch v 1.7.1 with a CUDA version that works on your cluster. We used CUDA 11.0 in this repo:

conda install pytorch==1.7.1 torchvision==0.8.2 torchaudio==0.7.2 cudatoolkit=11.0 -c pytorch

Then, install the packages in requirements.txt via pip:

pip3 install -r requirements.txt

Initialization with Pretrained weights for Training

Create the ./pretrained directory.

mkdir ./pretrained

Download ImageNet pretrained weights of the Swin Transformer and official weights of the BERT (Pytorch, uncased version) into the ./pretrained folder.

Trained weights of CrossVLT for Testing

Create the ./checkpoints directory.

mkdir ./checkpoints

Download CrossVLT model weights into the ./checkpoints folder.

Datasets

Follow README.md in the ./refer directory to set up subdirectories and download annotations. Download 2014 Train images [83K/13GB] from COCO, and extract the downloaded train_2014.zip file to ./refer/data/images/mscoco/images.

Training

We use DistributedDataParallel from PyTorch. The CrossVLT were trained using 2 x 24G RTX3090 cards. To run on multi GPUs (2 GPUs is used in this example) on a single node:

mkdir ./models

CUDA_VISIBLE_DEVICES=0,1 python3 -m torch.distributed.launch --nproc_per_node 2 --master_port 12345 train.py --dataset refcoco --swin_type base --lr 0.00003 --epochs 40 --img_size 480 2>&1 | tee ./models/refcoco

CUDA_VISIBLE_DEVICES=0,1 python3 -m torch.distributed.launch --nproc_per_node 2 --master_port 12345 train.py --dataset refcoco+ --swin_type base --lr 0.00003 --epochs 40 --img_size 480 2>&1 | tee ./models/refcoco+

CUDA_VISIBLE_DEVICES=0,1 python3 -m torch.distributed.launch --nproc_per_node 2 --master_port 12345 train.py --dataset refcocog --splitBy umd --swin_type base --lr 0.00003 --epochs 40 --img_size 480 2>&1 | tee ./models/refcocog

To store the training logs, we need to manually create the ./models directory via mkdir before running train.py.

• --dataset is the dataset name. One can choose from refcoco, refcoco+, and refcocog.
• --splitBy needs to be specified if and only if the dataset is G-Ref (which is also called RefCOCOg).
• --swin_type specifies the version of the Swin Transformer. One can choose from tiny, small, base, and large. The default is base.

Testing

To evaluate, run one of:

CUDA_VISIBLE_DEVICES=0 python3 test.py --swin_type base --dataset refcoco --split val --resume ./checkpoints/best_refcoco.pth --workers 1 --img_size 480

CUDA_VISIBLE_DEVICES=0 python3 test.py --swin_type base --dataset refcoco+ --split val --resume ./checkpoints/best_refcoco+.pth --workers 1 --img_size 480

CUDA_VISIBLE_DEVICES=0 python3 test.py --swin_type base --dataset refcocog --splitBy umd --split val --resume ./checkpoints/best_refcocog.pth --workers 1 --img_size 480

• --split is the subset to evaluate. One can choose from val, testA, and testB for RefCOCO/RefCOCO+, and val and test for G-Ref (RefCOCOg).
• --resume is the path to the weights of a trained CrossVLT.

Results

The complete evaluation results of the proposed model are summarized as follows:

Visualization

Comparison with ablated models (without both components (Basic model) / without cross-aware early fusion / without feature-based alignment) :

Comparison with previous state-of-the-art models (late fusion model (VLT) / vision-only early fusion model (LAVT)) :

@article{cho2023cross,
  title={Cross-aware Early Fusion with Stage-divided Vision and Language Transformer Encoders for Referring Image Segmentation},
  author={Cho, Yubin and Yu, Hyunwoo and Kang, Suk-Ju},
  journal={IEEE Transactions on Multimedia},
  year={2023},
  publisher={IEEE}
}