InfraSegFM/InfraSegDB research data for "Transferable Civil Infrastructure Defect Segmentation with Source-Aware Benchmarking and Routed SAM Adaptation"

This record contains the derived research data and reproducibility artifacts supporting the manuscript “Transferable Civil Infrastructure Defect Segmentation with Source-Aware Benchmarking and Routed SAM Adaptation.” The study introduces InfraSegDB, a source-aware benchmark containing 179,094 image-mask pairs from 40 source datasets and six acquisition platforms, and InfraSegFM, a parameter-efficient SAM ViT-B adaptation framework for transferable civil infrastructure defect segmentation. The deposit contains the final per-image and aggregate evaluation results used for the in-distribution, CrossSite, CrossTask, RealWorld, and limited-target-data evaluations at 0%, 5%, 15%, 25%, 50%, and 100% target-data ratios. It includes component-ablation results, the exact numerical source data and statistical-analysis outputs for Figs. 9–16, a non-pixel qualitative sample manifest with stable sample identifiers and integrity hashes, NVIDIA Jetson Orin NX profiling data and formal device logs, data dictionaries, provenance metadata, an analysis/visualization code snapshot, and SHA-256 checksums. Statistical comparisons use two-sided paired Wilcoxon signed-rank tests on matched task-sample identifiers, with Holm correction within each prespecified figure family. All model values in the deposit are stored as final canonical observations; no model-specific rescaling is performed at the figure stage. For CrossTask, Shared is the displayed and statistical reference, while Full remains the proposed routed model. Raw third-party images, original annotations, and qualitative triplet pixels are not redistributed because they remain governed by the licenses and terms of their respective providers. DATASET_SOURCES.csv provides the 40 source-family identifiers together with the access and source notes recovered from the project inventory. This deposit does not relicense third-party content. Third-party SAM backbone weights are not included. The maintained source code, trained InfraSegFM ViT-B checkpoint, paper-default configurations, and model card are available at https://github.com/YuyangBaoo/InfraSegFM. The synthetic demo_data directory in that repository is intended only for software verification and was not used for the reported results. Protocol version: InfraSegFM-paper-20260722-final.