MICCAI Challenge 2019 Dataset

About

There are 3 datasets (which are all from different pigs) each containing 5 keyframes. Keyframes are all single captures with a unique positioning of the endoscope and several positions of the structured light illuminator. The keyframes (with the exception of the last keyframe in each dataset) have additional data from interpolation using the camera kinematics to 'warp' the depth images to the new camera coordinates. These warped depth maps could be computed yourself manually since we provide the camera transforms but we are providing them for your convenience. Please do not assume they are perfect ground truth, due to errors in the kinematics or synchronization issues between the camera and video stream, there will be small errors in the warped vertex positions.

Each keyframe should contain the following files

camera_calibration.yaml This is an approximate calibration of the endoscope and may be inaccurate. It is in OpenCV format and you can load this data directly using OpenCV FileStorage class. However, depth estimation methods are expected to be robust to errors in camera calibration.
Left_Image.png This is the left camera view used when the structured light patterns for keyframe N were captured.
Right_Image.png This is the right camera view.
left_depth_map.tiff This is the point cloud as seen by the left camera. Each pixel in the .tiff file contains an (X,Y,Z) coordinate which is the vertex position in left camera space that the pixel projects to.
right_depth_map.tiff This is the point cloud as seen by the right camera.
point_cloud.obj This is a 3D vertex model you can load into software such as blender or meshlab to view the coordinates.
data/rgb.mp4 This is the sequence of video frames (stacked left above right) that we captured after the structured light sequence for keyframe N. You should be able to open and decode this file with either OpenCV or ffmpeg. Standard video players (vlc/quicktime/wmp) may not work.
data/frame_data.tar.gz This is a archive of frames with the naming format frame_data%06d.json. Each frame_data corresponds to a single video frame where the number in the file name can be used for matching. The frame data consists of the camera transform relative to the camera position where the keyframe illuminator image was captured (Left_Image.png/Right_Image.png) and the camera calibration data for that frame.
data/scene_points.tar.gz This is a archive of .tiff files with the naming format scene_points%06d.tiff. Like the frame_data files above, each scene_points corresponds to a single video frame where the number in the file name can be used for matching. The .tiff files contain the vertex coordinates warped from the keyframe illuminator image. Please note that as the camera motion introduces new pixels which are not captured in the illuminator scene, these pixels do not have an associated ground truth depth. You can find these pixels as they contain the value set (0,0,0) in the .tiff file.

Helpful Notes

Opening .tiff files can be performed with the Python library tifffile which is installable with pip. This format allow the storage of floating point images.
We apologize that interpolation is missing for dataset_1/keyframe_4 is missing. This was due to a logging error.