VoroCrack3d: An annotated data set of 3d CT concrete images with synthetic crack structures

VoroCrack3d is an annotated data set of 3d CT images of concrete with synthetic crack structures. Its main purpose is the training and testing of machine learning models for 3d crack segmentation. The data set comprises 1344 images together with their corresponding ground truths. The concrete backgrounds are cropped out sections of size 400x400x400 voxels of CT images of concrete. To this end, several different concrete samples were scanned (normal concrete (NC), high-performance concrete (HPC), ultra-high-performance concrete (UHPC), air pore concrete; without and with reinforcements (straight steel fibers, crimped steel fibers, hooked-end steel fibers, polypropylene fibers, fibers made of glass fiber-reinforced polymer). The original concrete images have a resolution between 2.8 and 106 micrometers.

The crack structures are modeled via minimum-weight surfaces in Voronoi diagrams according to the paper

[1] C. Jung, C. Redenbach, Crack Modeling via Minimum-Weight Surfaces in 3d Voronoi Diagrams, Journal of Mathematics in Industry, 13, 10 (2023). https://doi.org/10.1186/s13362-023-00138-1.

The surfaces are discretized, dilated and superimposed on the concrete backgrounds.

The data set offers a high variety regarding concrete types, noise levels and crack widths, shapes, regularity and branching. This makes it suitable for studying the generalizability and robustness of 3d crack segmentation methods.

______________________________________________________________________________________________

The folder 'data' contains seven subfolders, each containing the data generated from a specific concrete type (NC, HPC, air pore concrete, polypropylene fiber-reinforced concrete, steel fiber-reinforced concrete (straight, crimped and hooked-end steel fibers)).

Each subfolder again contains four subfolders according to the point process model that was used for generating the 3d Voronoi diagrams. The point processes and Voronoi diagrams are restricted to windows of size 400x150x400. 

- 'hc': Hard core point process with 60% volume density and intensity 0.000025 obtained from force-biased sphere packing.
- 'matclust': Matérn cluster process with parent intensity 0.0002/50, offspring intensity 50 and cluster radius 20.
- 'ppp': Poisson point process with intensity 0.0002.
- 'ppp-scaled': Poisson point process with intensity 0.0002 (but inside 200x150x200 window). The resulting Voronoi diagram is stretched in x- and z- direction by a factor of 2.

Each of these contains five subfolders: one for the 3d input images, two for the corresponding labels (ground truths; one with and one without pores/fibers), one for the input and label previews (slice z=200 for each of the images) and a misc folder containing the concrete background without crack and, if applicable, the pore/fiber segmentation image.

The data itself then contains 48 images:
1a-1d: crack with up to seven branches; fixed crack width (~1 voxel).
2a-2d: crack with up to four branches; fixed crack width (~1 voxel).
3a-3d: crack with up to one branch; fixed crack width (~1 voxel).
4a-4d: crack with no branches; fixed crack width (~1 voxel).
5a-5d: crack with no branches; fixed crack width (~3 voxels).
6a-6d: crack with no branches; fixed crack width (~5 voxels).
7a-7d: crack with no branches; fixed crack width (~7 voxels).
8a-8d: crack with up to seven branches; multiscale crack (bernoulli parameter 0.01);
9a-9d: crack with up to seven branches; multiscale crack (bernoulli parameter 0.02);
10a-10d: crack with up to seven branches; multiscale crack (bernoulli parameter 0.05);
11a-11d: crack with up to seven branches; multiscale crack (bernoulli parameter 0.1);
12a-12d: crack with up to seven branches; multiscale crack (bernoulli parameter 0.2);

The names 'a'-'d' indicate level of added noise added to the image:
a: None.
b: Uniformly on [-sigma,sigma] 
c: Uniformly on [-2*sigma,2*sigma] 
d: Uniformly on [-4*sigma,4*sigma] 
Negative values are mapped to 0. 
For inputs of type int, noise values are rounded to the nearest integer.
(sigma = standard deviation of voxel greyvalues in image)

Note that the grey values in the ground truths correspond to the local crack width. They can be thresholded to obtain binary masks.

For more details, we refer to [1].