This repository contains Unofficial code to help you generate additional data, such as disparity, depthmap and flow samples in the left rectified frame of reference. Additionally tools are provided to generate the interpolated ground truth sequences based on keyframe ground truth data.
If you are using the Stereo correspondence and reconstruction of endoscopic data cite the challenge's paper
@article{allan2021stereo,
title={Stereo correspondence and reconstruction of endoscopic data challenge},
author={Allan, Max and Mcleod, Jonathan and Wang, Congcong and Rosenthal, Jean Claude and Hu, Zhenglei and Gard, Niklas and Eisert, Peter and Fu, Ke Xue and Zeffiro, Trevor and Xia, Wenyao and others},
journal={arXiv preprint arXiv:2101.01133},
year={2021}
}
This repository contains code developed as a part of the paper MSDESIS: Multi-task stereo disparity estimation and surgical instrument segmentation. If you end up using code provided in this repository, please consider citing
@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 dataset was made publicly available after the completion of the SCARED challenge. To get access you need to go to the SCARED challenge download webpage, create a user account in the grand-challenge platform, join the challenge and follow the provided instructions.
We've established a data format to facilitate development. Depth and disparity data are loaded and manipulated as floats. If disparity or depth information is not available for a specific pixel, its values are represented by nan.
The .obj pointclouds, provided with every keyframe, contain HxW points, with H, W the height and with of the monocular frame. Since SCARED does not provide full coverage, some of those vertices are represented as nan. Our loading functions remove such points which results to much smaller pointclouds containing only points for which we know ground truth information. Although not used by our scripts, we provide code to save pointclouds as ply. In that case we save only points with known ground truth.
The provided .tiff keyframe files encode unknown values as nan, whereas the provided interpolated .tiff files in the sequences zeros(we haven't check every sequence). Since we want all unknown points to have nan values, when loading, our functions replace 0 vectors with nan values.
To facilitate sample preview, we store both generated disparity and depthmaps as 16bit uint pngs. All depth values are in mm distance and disparity is measured as the difference in y directions between the coordinates of a point in the left stereo rectified image with its corresponding point in the right stereo rectified image. In order to maintain decimal information when storing samples as .png, we scale the disparity and depth values by a configurable argument called scale_factor(default is 256.0) This maps a range of 0-255 to 0-65280 and then store them as 16bit unsigned integers. If there is a need of storing values greater than 255, one can adjust the the scale factor to something that will cast the sample range to span values 0-2^16. Nan values are stored as 0. When loading such samples the scale_factor is used to remap 0-2^16 values to the correct range and 0 values are replaced by nan. This process is obviously lossy but it maintains correct information up to 2 decimal points when a scale_factor of 256.0 is used.
In addtition to hight level data extraction and disparity generation scripts
the repository, provides python code to load and save samples provided with
the original dataset as well as functions to store and load depthmap and
disparity samples with decimal information encoded in 16-bit uint .png.
It also includes code to manipulate samples and create additional data,
such as disparity samples. Included functions are able to generate depthmaps,
3D images, disparities and pointclouds from any of the aforementioned domains.
Still if you are to use the the provided functions and not the scripts you need
to check the validity of the outcome. For instace, ptcloud_to_disparity() can
generate a disparity image based on a pointcloud, the result is meaningless if
the provided pointcloud is not rotated to the rectified frame of reference and
the Projection matrices are not obtained from the stereo rectification process.
This project was build using anaconda. Assuming that anaconda is already installed in the target machine, a anaconda environment suitable to run this code can be created using the following steps.
- navigate to this project's folder
- create an environments (e.g. scared_toolkit) using the provided requirements.txt
conda create --name scared_toolkit --file requirements.txtactivate the anaconda environment
conda activate scared_toolkitAll the scripts listed bellow expect an standard file structure. Before using ensure that the initial scared dataset follows the file structure described bellow
.
├── dataset_1 # each subdataset folder should follow the dataset_{$dataset_number} notation
│ ├── keyframe_1 # each keyframe folder should follow the keyframe_{$keyframe_number} notation
│ │ │ endoscope_calibration.yaml
│ │ │ left_depth_map.tiff
│ │ │ Left_Image.png
│ │ │ point_cloud.obj
│ │ │ right_depth_map.tiff
│ │ │ Right_Image.png
│ │ └── data
│ │ ├── frame_data.tar.gz
│ │ ├── rgb.mp4
│ │ └── scene_points.tar.gz
: : :
│ └── keyframe_M
│ └── data
: :
└── dataset_N
├── keyframe_1
: :
└── keyframe_M
This script does the following:
- splits the rgb.mp4 into left and right .png images
- unpacks and the contents of the scene_points.tar.gz and splits them to left and right
- reads the endoscope calibtration.yaml, port it to opencv format and add stereo rectification related parameters
- [optional] undistort the rgb frames and depthmaps
--undistort - [optional] generate depthmap in .png format
--depth - [optional] stereo rectify the rgb images from the rgb.mp4
--disparity - [optional] stereo rectify depthmaps
--disparity - [optional] generate left disparity maps
--disparity
python -m scripts.extract_sequence_dataset root_dir [--out_dir] [--recursive] [--depth] [--undistort] [--disparity] [--alpha] [--scale_factor]root_dir root directory under which keyframe data are stored
--out_dir where to store the resulting dataset, if not set, generated files will be stored in src folders
--recursive scans for keyframe_* directories under root_dir and processes them all
--depthgenerate_depthmap in the original frame of reference (.pngs)
--undistortgenerate undistorted depthmap and left rgb in the original frame of reference
--disparitygenerate rectified views and disparity maps
--alpha corresponds to the alpha rectification parameters used in the OpenCV stereo rectification function
--scale_factor refer to this
){kind=link}
This scripts offers the same functionality as extract_sequence_dataset but with the exception that it generates a smaller dataset only using the keyframes, completely ignoring the rgb.mp4 and scene_points.tat.gz sequences. Additionally it offers the ability to overwrite the provided .obj groundtruth pointcloud.
python -m scripts.generate_keyframe_dataset root_dir [--out_dir] [--recursive] [--depth] [--undistort] [--disparity] [--pt_cloud] [--alpha] [--scale_factor]--ptcloud name of the pointcloud to provide reference, .ply are supported, must be placed inside keyframe dirs.
This scripts offers the same functionality as extract_sequence_dataset but with the exception that it generates the ground truth interpolated sequence based on the point_cloud.obj( can be overwritten with a .ply file) and the frame_data.tar.gz endoscope pose sequence. This has the advantage of completely removing the overall size of the dataset the scene_points.tat.gz files from the dataset reducing its size to only 7.2GB and making it portable. The generated files are not the exact same as the contents of the scene_points.gz files possibly due to numerical precision. For this reason the use of this script should be limited to only the training datasets and not the test datasets.
python -m scripts.generate_sequence_dataset root_dir [--out_dir] [--recursive] [--depth] [--undistort] [--disparity] [--pt_cloud] [--alpha] [--scale_factor]This scripts can be used to generate flow maps and store them in the same format used by kitti. The command line interface is similar to the previous scripts but only supports flow generation in the original frame of reference.
python -m scripts.generate_flow_sequence root_dir [--recursive] [--out_dir] [--ptcloud]