southflame/restr

Official repository of the "ReSTR: Convolution-Free Referring Image Segmentation Using Transformers (CVPR'22)"

ReSTR: Convolution-Free Referring Image Segmentation Using Transformers

This repository contains the official source code for our paper:
ReSTR: Convolution-Free Referring Image Segmentation Using Transformers
Namyup Kim¹, Dongwon Kim¹, Cuiling Lan², Wenjun Zeng², and Suha Kwak^1
1POSTECH CSE, ²Microsoft Research Asia
CVPR, 2022.

Environment setup

• Python 3.7.13
• PyTorch 1.13.1+cu117

Instructions:

conda env create -f restr.yaml
conda activate restr

Data Setup

1. Setting

• Download or use symlink, such that the MS COCO images are under data/coco/images/train2014/
• Download or use symlink, such that the ReferItGame data are under data/referit/images and data/referit/mask

• Download, git clone, or use symlink, such that refer is under external. Then strictly follow the Setup and Download section of its README. Also, put the refer folder in PYTHONPATH as export PYTHONPATH=${PYTHONPATH}:/my/restr/path/external/refer
• Download, git clone, or use symlink, such that the MS COCO API is under external (i.e. external/coco/PythonAPI/pycocotools)

2. Data preparation

python build_batches.py -d Gref -t train --img-size 480
python build_batches.py -d Gref -t val --img-size 480
python build_batches.py -d unc -t train --img-size 480
python build_batches.py -d unc -t val --img-size 480
python build_batches.py -d unc -t testA --img-size 480
python build_batches.py -d unc -t testB --img-size 480
python build_batches.py -d unc+ -t train --img-size 480
python build_batches.py -d unc+ -t val --img-size 480
python build_batches.py -d unc+ -t testA --img-size 480
python build_batches.py -d unc+ -t testB --img-size 480
python build_batches.py -d referit -t trainval --img-size 480
python build_batches.py -d referit -t test --img-size 480

3. Directory Structure After Sutup and Data Preparation

├─ ./data              
      ├─ mscoco   
      │   ├─ Gref_480_batch
      │   │   ├─ train_batch
      │   │   |     ├─ Gref_train_0.npz
      │   │   |     ├─ Gref_train_1.npz
      │   │   |     └─ ...
      |   |   ├─ train_image
      │   │   ├─ train_label 
      │   │   ├─ val_batch
      │   │   ├─ val_image
      │   │   └─ val_label
      │   ├─ unc_480_batch
      │   └─ unc+_480_batch
      ├─ referit
      │   └─ referit_480_batch
      │       ├─ trainval_batch
      │       └─ text_batch
      ├─ Gref_emb.npy
      ├─ referit_emb.npy
      ├─ vocabulary_Gref.txt
      └─ vocabulary_referit.txt

Training

python train_restr.py --data_dir ./data/mscoco/Gref_480_batch --adamW
python train_restr.py --data_dir ./data/mscoco/unc_480_batch --adamW
python train_restr.py --data_dir ./data/mscoco/unc+_480_batch --adamW
python train_restr.py --data_dir ./data/referit/referit_480_batch --set trainval --valset test --adamW

Evaluation

cd eval
python evaluate.py --data_dir ../data/mscoco/Gref_batch --restore_refseg ../weights/test --set val --iters 25000 --input-size 480,480 --is_vis

Citation

@inproceedings{kim2022restr,
  title={Restr: Convolution-free referring image segmentation using transformers},
  author={Kim, Namyup and Kim, Dongwon and Lan, Cuiling and Zeng, Wenjun and Kwak, Suha},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={18145--18154},
  year={2022}
}

Acknowledgement

This code is built upon the following public repositories.

• https://github.com/chenxi116/TF-phrasecut-public
• https://github.com/rstrudel/segmenter