Skip to content

Latest commit

 

History

History
30 lines (23 loc) · 1.72 KB

README.md

File metadata and controls

30 lines (23 loc) · 1.72 KB
Raw

Datasets

Here we describe the steps for using the Endoscopic Vision 2017 [1] and 2018 [2] for instrument-type segmentation.

Downloading the data

  1. Download the 2017 dataset from the challenge webpage here.
  2. Download the 2018 dataset with our instrument-type annotations here.

Organizing the data

Organize the data of both datasets into the appropriate splits. Note for training on all the data, use the 2018 labels compatible with the 2017 dataset (final_labels_2017.json).

python organize2017.py --data-dir /path/to/original/data \
--save-dir /path/to/save/organized/data --cropped
python organize2018.py --data-dir /path/to/data/with/our/annotations \
--save-dir /path/to/save/organized/data --label-path /path/to/labels.json

Afterwards, convert the dataset to the MS-COCO format. Required for the temporal consistency module and using our pre-trained weights that employ the Mask R-CNN Detectron v.1 code.

python robotseg_to_coco.py --image-dir /path/to/images \
    --anns-dir /path/to/binary_annotations --save-dir /path/to/save/coco/annotations \
    --dataset <dataset_name> --group-name <dataset_split> \
    --json-prefix <prefix_for_json_file>

References

[1] Allan, M., Shvets, A., Kurmann, T., Zhang, Z., Duggal, R., Su, Y.H., , et al.: 2017 robotic instrument segmentation challenge. arXiv preprint arXiv:1902.06426 (2019) [2] Allan, M., Kondo, S., Bodenstedt, S., Leger, S., Kadkhodamohammadi, R., Luengo, I., Fuentes, F., Flouty, E., Mohammed, A., Pedersen, M., et al.: 2018 robotic scene segmentation challenge. arXiv preprint arXiv:2001.11190 (2020)