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. |
 |  |
 |  |
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";
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}
}