ISTR: End-to-End Instance Segmentation with Transformers (https://arxiv.org/abs/2105.00637)

This is the project page for the paper:

ISTR: End-to-End Instance Segmentation via Transformers,
Jie Hu, Liujuan Cao, Yao Lu, ShengChuan Zhang, Yan Wang, Ke Li, Feiyue Huang, Ling Shao, Rongrong Ji,
arXiv 2105.00637

⭐ Highlights:

• GPU Friendly: Four 1080Ti/2080Ti GPUs can handle the training for R50, R101 backbones with ISTR.
• High Performance: On COCO test-dev, ISTR-R50-3x gets 46.8/38.6 box/mask AP, and ISTR-R101-3x gets 48.1/39.9 box/mask AP.

Updates

• (2021.05.03) The project page for ISTR is avaliable.

Models

• The inference time is evaluated with a single 2080Ti GPU.
• We use the models pre-trained on ImageNet using torchvision. The ImageNet pre-trained ResNet-101 backbone is obtained from SparseR-CNN.

Installation

The codes are built on top of Detectron2, SparseR-CNN, and AdelaiDet.

Requirements

• Python=3.8
• PyTorch=1.6.0, torchvision=0.7.0, cudatoolkit=10.1
• OpenCV for visualization

Steps

1. Install the repository (we recommend to use Anaconda for installation.)

conda create -n ISTR python=3.8 -y
conda activate ISTR
conda install pytorch==1.6.0 torchvision==0.7.0 cudatoolkit=10.1 -c pytorch
pip install opencv-python
pip install scipy
pip install shapely
git clone https://github.com/hujiecpp/ISTR.git
cd ISTR
python setup.py build develop

2. Link coco dataset path

ln -s /coco_dataset_path/coco ./datasets

3. Train ISTR (e.g., with ResNet50 backbone)

python projects/ISTR/train_net.py --num-gpus 4 --config-file projects/ISTR/configs/ISTR-R50-3x.yaml

4. Evaluate ISTR (e.g., with ResNet50 backbone)

python projects/ISTR/train_net.py --num-gpus 4 --config-file projects/ISTR/configs/ISTR-R50-3x.yaml --eval-only MODEL.WEIGHTS ./output/model_final.pth

5. Visualize the detection and segmentation results (e.g., with ResNet50 backbone)

python demo/demo.py --config-file projects/ISTR/configs/ISTR-R50-3x.yaml --input input1.jpg --output ./output --confidence-threshold 0.4 --opts MODEL.WEIGHTS ./output/model_final.pth

Citation

If our paper helps your research, please cite it in your publications:

@article{hu2021ISTR,
  title={ISTR: End-to-End Instance Segmentation via Transformers},
  author={Hu, Jie and Cao, Liujuan and Lu, Yao and Zhang, ShengChuan and Li, Ke and Huang, Feiyue and Shao, Ling and Ji, Rongrong},
  journal={arXiv preprint arXiv:2105.00637},
  year={2021}
}