This repository contains unofficial code to assist with the use of the RIS2017 in stereo setting. Additionally this repository can be used to pre-process the monocular RIS2017 data by cropping and de-interlacing the provided frames.

If you are using the Robotics instrument Segmentation 2017 Endovis challenge dataset cite the official paper

@article{allan20192017,
  title={2017 robotic instrument segmentation challenge},
  author={Allan, Max and Shvets, Alex and Kurmann, Thomas and Zhang, Zichen and Duggal, Rahul and Su, Yun-Hsuan and Rieke, Nicola and Laina, Iro and Kalavakonda, Niveditha and Bodenstedt, Sebastian and others},
  journal={arXiv preprint arXiv:1902.06426},
  year={2019}
}

If this repository contains code useful to your research consider citing MSDESIS: Multi-task stereo disparity estimation and surgical instrument segmentation from which this code originates

@ARTICLE{9791423,
  author={Psychogyios, Dimitrios and Mazomenos, Evangelos and Vasconcelos, Francisco and Stoyanov, Danail},
  journal={IEEE Transactions on Medical Imaging}, 
  title={MSDESIS: Multi-task stereo disparity estimation and surgical instrument segmentation}, 
  year={2022},
  volume={},
  number={},
  pages={1-1},
  doi={10.1109/TMI.2022.3181229}}

The project provides code to manipulate RIS 2017 data and supports:

• Converting the provided calibration files to OpenCV format
• Stereo Rectification of images.
• De-interlacing
• Binary mask generation.
• Evaluation script for binary segmentation.

The dataset had been made publicly available after the completion of the EndoVis challenge. To get access you need to go to the RIS2017 EndoVis challenge download webpage, create a user account in the grand-challenge platform, join the challenge and follow the provided instructions on how to download the dataset.

The table below, lists color statistics and binary segmentation pixel coverage per dataset. Color statistics were computed only using the left frame from each sample, after de-interlacing and cropping it. The range is normalized from 0-255 to 0-1.

• Extrinsic Rotation matrix

The provided calibration includes a three value entry named Extrinsic-Omega which describes to the rotation from left to right camera frame. Assuming that the rotation is expressed using the Rodrigues' formula, the scripts convert 3x3 Extrinsic-Omega to a rotation matrix which are later used for stereo rectification

Based on the provided calibration parameters, albeit the frame size is indicated to be 1280x1024, the provided samples are of size 1920x1080 with the rgb image places in the middle of the samples and black borders around it. Center cropping the frames to 1280x1024, results to images with back borders. To remove the borders completely and to make use of the provided calibration we crop frames to 1263x1009 starting from pixel 328,37.

Sample are extracted from interlaced video. This results to artifacts whenever an object is moving fast between frames. Because the full video sequence is not provided, we are limited to simple de-interlacing techniques where we delete odd image rows and interpolate their values based color information of the even lines.

The left and right view are not perfectly synchronized in time. This can be easily seen in views where a tool is moving fast in the vertical direction. Dataset 1- Frame 207

Using the provided calibration files to stereo rectify the dataset we find the the following:

• Datasets 3,6 exhibit vertical disparity of around 20 pixels after stereo rectification.
• Datasets 5,7,8,9,10 exhibit vertical disparity around 5 pixels after stereo rectification.
• Datasets 1,2 and 4 do not exhibit visible vertical disparity after stereo rectification.

This project was build using anaconda. Assuming that anaconda is already installed create an environment by running:

conda create --name ris_toolkit --file requirements.txt
conda activate ris_toolkit

After downloading the data, you need to re-arrange the files such that they follow the file structure below.

.
├── train_set                       # folder containing train samples
│   └── instrument_dataset_i        # i corresponds to dataset number (0-8)
│       ├── left_frame              # dir containing left frames
│       ├── right_frame             # dir containing right frames
│       ├── ground_truth 
│       │   └── label_1             # dir contains ground truth labels for tool 1
│       │   └── label_n             # dir contains ground truth labels for tool n
│       ├── mappings.json
│       └── camera_calibration.txt  # 12 line calibration file.
└── test_set                        # folder containing test samples
    └── instrument_dataset_i        # i corresponds to dataset number (0-10)
        ├── left_frame              # dir containing left frames
        ├── right_frame             # dir containing right frames
        ├── ground_truth 
        │   └── label_1             # dir contains ground truth labels for tool 1
        │   └── label_n             # dir contains ground truth labels for tool n
        ├── mappings.json
        └── camera_calibration.txt  # 12 line calibration file.

The following scripts generate a processed version of the RIS_2017 dataset following the above file structure. The stereo rectified frames new dataset will be de-interlaced, cropped and interpolated to 1280x1024. Additionally a stereo_calib.json, OpenCV compatible, calibration file will be generated containing both the provided calibration parameters and the stereo rectification matrices in OpenCV format. The rectification cropping can be adjusted using the --rect_alpha flag.

To create the a stereo rectified dataset de-interlaced and cropped to 1280x1024 dataset,use the following script for the test and training sequence respectively

python -m scripts.generate_stereo_binary_dataset /path_to_original/train_set /path_to_store_the_new/train_set --alpha_rect -1
python -m scripts.generate_stereo_binary_dataset /path_to_original/test_set /path_to_store_the_new/test_set --alpha_rect -1

To create the a monocular de-interlaced and cropped to 1280x1024 dataset, use the following script for the test and training sequence respectively

python -m scripts.generate_clean_monocular_dataset /path_to_original/train_set /path_to_store_the_new/train_set 
python -m scripts.generate_clean_monocular_dataset /path_to_original/test_set /path_to_store_the_new/test_set 

To generate a csv file containing a list of all left, right images and binary segmentation masks under a specific folder recursively you can use the following script

• The --binary flag dictates if the outcome .csv file will contain ground truth binary segmentation masks as the third column.

python -m scripts.generate_io_csv.py path/to/dataset ./path/to/write/csv [--binary]

You can use this script to validate predictions using IoU metric. The script will compute a score fore every dataset in the test sequence

python -m scripts.evaluate  --gt_dir /path/to/test/dataset/root/dir --algorithm_output_dir /path/to directory/samples/for/eval/are/stored