Transparent Transformer Segmentation

Introduction

This repository contains the data and code for IJCAI 2021 paper Segmenting transparent object in the wild with transformer.

Environments

• python 3
• torch = 1.4.0
• torchvision
• pyyaml
• Pillow
• numpy

INSTALL

python setup.py develop --user

Data Preparation

1. create dirs './datasets/transparent/Trans10K_v2'
2. put the train/validation/test data under './datasets/transparent/Trans10K_v2'. Data Structure is shown below.

Trans10K_v2
├── test
│   ├── images
│   └── masks_12
├── train
│   ├── images
│   └── masks_12
└── validation
    ├── images
    └── masks_12

Download Dataset: Google Drive. Baidu Drive. code: oqms

Network Define

The code of Network pipeline is in segmentron/models/trans2seg.py.

The code of Transformer Encoder-Decoder is in segmentron/modules/transformer.py.

Train

Our experiments are based on one machine with 8 V100 GPUs with 32g memory, about 1 hour training time.

bash tools/dist_train.sh $CONFIG-FILE $GPUS

For example:

bash tools/dist_train.sh configs/trans10kv2/trans2seg/trans2seg_medium.yaml 8

Test

bash tools/dist_train.sh $CONFIG-FILE $GPUS --test TEST.TEST_MODEL_PATH $MODEL_PATH

Citations

Please consider citing our paper in your publications if the project helps your research. BibTeX reference is as follows.

@article{xie2021segmenting,
  title={Segmenting transparent object in the wild with transformer},
  author={Xie, Enze and Wang, Wenjia and Wang, Wenhai and Sun, Peize and Xu, Hang and Liang, Ding and Luo, Ping},
  journal={arXiv preprint arXiv:2101.08461},
  year={2021}
}