PoreBoostGAN
Description
The MRCCM dataset, accessible via the Digital Rocks Portal(https://www.digitalrocksportal.org/projects/362), comprises multi-resolution digital rock scans of two distinct carbonate samples, each imaged at two different resolutions. This open-access dataset is particularly valuable for advancing research in digital rock analysis, including the study of partial volume effects, segmentation techniques, and the benchmarking of deep learning models for super-resolution and segmentation tasks. We specifically selected two digital rocks from the MRCCM dataset for validation purposes. To ensure compatibility with the Biogenic dataset—which features scans of more complex and long-term geological structures—the MRCCM dataset has been converted to .tif format. The Biogenic dataset includes three in-situ resolution scans, enabling direct validation of ultra-high-resolution models against ground truth data. In contrast, the MRCCM dataset, with only two available resolutions, has undergone a downsampling strategy to generate the necessary data for comprehensive training and evaluation. The SR folder within the dataset is particularly noteworthy. It contains data specifically prepared for testing the extrapolation capabilities of super-resolution models, making it an essential resource for researchers focused on developing and refining these models. The root directory of the dataset includes 3D reconstruction training sets for digital rock cores, while the SR folder provides datasets tailored for evaluating the performance of super-resolution techniques in extrapolating higher-resolution details. This structure ensures that researchers have access to both the foundational data necessary for training and specialized datasets designed for rigorous testing of super-resolution methods.
Files
Steps to reproduce
All 3D images are compatible with ImageJ for viewing and analysis. Preprocessing scripts and super-resolution code associated with this dataset will be publicly released upon the publication of the related research paper.