Spall-to-Full: Datasets and Model Checkpoints for Fire-Induced Concrete Spalling Depth Completion

This record contains the datasets and trained model checkpoints associated with the manuscript “Semantics Guided RGB-D Quantification of Fire-induced Concrete Spalling”, which is currently under peer review. The manuscript presents Spall-to-Full, a semantics-guided RGB-D depth-completion framework developed to restore incomplete metric depth maps of concrete-spalling scenes and support subsequent geometric damage quantification. The released material includes fire-induced explosive-spalling data and additional load-induced spalling and real-world tunnel scenes. The record contains: data.zip: the training and validation dataset of Spall to Full, comprising 691 high-resolution RGB-D images, including 553 training images and 138 validation images. Each subset contains RGB images, metric depth maps, semantic label masks, and one-hot semantic masks. All images are at a resolution of 3840×2160 or 4096×3072, and the depth maps have been registered to the RGB images. Test Dataset.zip: the quantitative test dataset containing RGB images, ground-truth metric depth maps, semantic masks, ROI masks, and sparse-depth inputs generated by removing 40%, 60%, and 80% of the valid depth pixels. All images are at a resolution of 4096×3072, and the depth maps have been registered to the RGB images. checkpoint.zip: the trained Spall-to-Full model checkpoint files used for inference and quantitative evaluation.Three models with different input resolutions are provided, specifically sizes 1456, 1036, and 518. Real world Tunnel Test.zip: the real-world tunnel evaluation dataset acquired in China, including challenging upward and highly oblique observations under substantial domain differences. The released depth data use millimetres as the metric unit. The corresponding source code, environment requirements, training and evaluation scripts, directory conventions, and usage instructions are available at: https://github.com/MikeyDong/Spall-to-Full NOTE:The datasets and checkpoint files archived in this Zenodo record are the same as the corresponding resources distributed through the Baidu Netdisk links provided in the GitHub repository. The bibliographic information of the associated manuscript will be added after publication.