SIMUPOR: Benchmark for multi-phase segmentation of porous media tomography images
A benchmark dataset of 68 3D volumetric images of porous media with varying grain geometry and composition is composed. 3D volumetric images were obtained from the experiments conducted in (Al-Raoush, 2014) to study the effect of grain geometry on the morphology of non-aqueous phase liquids in porous media. The dataset used in this study represent types of soil that are important to many applications in soil, earth and environmental sciences. While the contrasts between the phases in such images are good, the need to obtain an accurate and unsupervised segmentation is required. This is extremely critical when the problem at had involves computations of interfaces between different phases in the images such as interfacial areas and mass transfer computations. Moreover, experiments that deal with dynamic systems generate a very large data that requires an unsupervised segmentation algorithms for an efficient processing. The volumes in SIMUPOR dataset correspond to samples from 34 different experiments, each corresponding to a specific constitution of porous medium. Among the 68 volumetric images used, 40 belong to the experiments that employ silica sand to model the porous media whereas the remaining 28 used quartz crystals. In addition to this variation in the shape of grains, there is also a variety in the size, with the median grain diameter ranging from 0.179 to 0.433 mm. This provides a comprehensive benchmark to check for robustness of any segmentation algorithms to changes in porous media composition.
Steps to reproduce
Dataset consists of 38 folders, named "Column_xx", each belonging to a different experiment. File structure for each of these folders and the variables contained in the .mat files is as follows: Column_xx --->B --->--->(top/bottom).mat --->--->--->[Variable] top/bottom => 256x256x256 Grayscale Volumetric stack corresponding to the top/bottom half of the scanned column. --->--->(top/bottom)_out.mat --->--->--->[Variable] SalMap => 256x256x256x4 tensor containing output of the proposed method at each of the 4 supervoxel resolutions. --->--->gt_(top/bottom).mat --->--->--->[Variable] gt_(top/bottom) => 256x256x256 ground truth for the top/bottom half of the column with regions of interest labelled from 0 to 3.