TUM RGB-D Scribble-based Segmentation Benchmark

Description

The RGB-D dataset contains the following

• The number of RGB-D images is 154, each with a corresponding scribble and a ground truth image.
• Every image has a resolution of 640 × 480 pixels.
• The measurement of the depth images is millimeter.
• The categorization differentiates between 95 classes.
• All scenes are indoor.

LabeledImages	This folder includes all images with the naming convention: [scene]_[number]_[image type].png, where scene is either bedroom, kitchen, livingroom or random and image type is either image, depth, scribbles or gt.
RawData	In this folder the original data in .xcf format can be found.
UnalignedDepth	One can find here all depth images before they were registered.
rgbd_palette.gpl	The ground truth and scribble images are converted to indexed mode. The related color palette is saved in this file.
LabelColorMapping.csv	This file describes which color belongs to which object class.
displayLabeledImages.py	For visualization this script provides an overview of one image with the associated classes.
Calibration	This folder contains the scripts, parameters and the images which were used for finding the parameters and for registering the depth images.

Example

For visualizing the point cloud, this matlab script can be used.

figure( 1, "visible", "off" );
depth = imread('LabeledImages/kitchen_22_depth.png');
depth = double(depth);
img = imread('LabeledImages/kitchen_22_image.png');
surf(depth, img, 'FaceColor', 'texturemap', 'EdgeColor', 'none' )
view(158, 38)
print -dpng pointCloud.png;
ans = "pointCloud.png";

Citation

If you use the dataset, please cite as following

@misc{tum-rgbd_scribble_dataset,
 author    = {Caner Hazirbas and Andreas Wiedemann and Robert Maier and Laura Leal-Taixé and Daniel Cremers},
 title     = {TUM RGB-D Scribble-based Segmentation Benchmark},
 howpublished = {\url{https://github.com/tum-vision/rgbd_scribble_benchmark}},
 year = {2018}
}