BCV-Uniandes / ISINet

Pytorch implementation of the MICCAI 2020 paper ISINet: An Instance-Based Approach for Surgical Instrument Segmentation.

ISINet

This is the Pytorch implementation of ISINet: An Instance-Based Approach for Surgical Instrument Segmentation published at MICCAI2020.

Installation

Requirements:

• Python >= 3.6
• Pytorch == 1.4
• numpy
• scikit-image
• tqdm
• scipy == 1.1
• flownet2
• Detectron v.1 (for using our pre-trained weights)

Pre-trained Weights

Pre-trained weights are publicly available on the project page.

Additional Annotations EndoVis 2018 Dataset

Additional annotations for the EndoVis 2018 Dataset are publicly available on the project page.

Data Preparation

Check the instructions detailed in data/README.md

Perform Inference

python -W ignore main.py --inference --model FlowNet2  --batch_size batch_size --number_workers num_workers \
--inference_dataset RobotsegTrackerDataset \ --inference_dataset_img_dir /path/to/images \ --inference_batch_size batch_size \
  --inference_dataset_coco_ann_path /path/to/coco/annotations/file.json \
  --inference_dataset_segm_path /path/to/mask-rcnn/inference/segm.json \
  --inference_dataset_ann_dir /path/to/annotations \
  --inference_dataset_cand_dir /path/to/save/candidates \ --inference_dataset_nms 'True' \
  --save /path/to/save/predictions \
  --inference_dataset_dataset '2017' or '2018' \
  --inference_dataset_maskrcnn_inference 'False' \
  --assignment_strategy 'weighted_mode' \ --inference_dataset_prev_frames 7 \
  --threshold 0.0 for 2017 and 0.5 for 2018 \
  --resume /path/to/flownet/checkpoint --num-classes number_of_classes

Reference

If you found our work useful in your research, please use the following BibTeX entry for citation:

@article{ISINet2020,
  title={ISINet: An Instance-Based Approach for Surgical Instrument Segmentation},
  author={Cristina Gonz{\'a}lez and Laura Bravo-S{\'a}nchez and Pablo Arbelaez},
  journal={arXiv preprint arXiv:2007.05533},
  year={2020}
}

Acknowledgements

Our code is build upon FlowNet2, we thank the authors for their contributions to the community.