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

Version 1.1.0 provides the numerical experimental results supporting “Transferable Civil Infrastructure Defect Segmentation with Source-Aware Benchmarking and Routed SAM Adaptation,” aligned with the revised manuscript and Supplementary Material. The release contains four independent ZIP archives and 48 CSV tables in total:01_baseline.zip — 6 tables of comparison-model results and recorded training costs.02_main_3runs.zip — 7 tables of Full, Shared and component-ablation results, including the task-to-asset and reported-run seed maps.03_inference_routing.zip — 26 tables of candidate selection, box-coordinate perturbation, full-image-box and empty-spatial-prompt evaluation, routing statistics and quality-head calibration.04_jetson_full98.zip — 9 tables of device measurements, model interfaces and fixed-image records.Each ZIP retains its corresponding numbered directory. The four archives form a complete, non-overlapping partition of the verified experimental CSV collection. The files include per-run, per-task and per-image records, condition summaries, learned-gate statistics, source-training timings and Jetson measurements. The reporting-seed crosswalk follows Supplementary Tables S25, S27 and S28; the task-to-asset map follows Table S26. Original execution identifiers remain traceable. Main comparison summaries use three independent training runs. Prompt studies retain their fixed model checkpoint and finalized quality-head profiles. Fixed candidate 0, original-head selection, finalized quality-head selection, predicted-IoU selection and offline same-candidate Oracle references retain their recorded definitions. Image-mean Dice averages per-image scores. Task-macro Dice first averages within each recorded task-context group, then gives those groups equal weight. A fixed full-image box remains a spatial prompt. Box-coordinate perturbation percentages and target-training-data percentages are separate variables. Device tables cover eight models in two precisions with 98 measured images per configuration. The four archives preserve the original CSV bytes and numerical precision. Their server MD5 checksums were verified against the local release files on 2 November 2026. No additional model training or inference was performed for this release. Archive SHA-256 checksums:01_baseline.zip: ddeef561ac34c928a0a838d8e220aeaab6a81e872a3a549b6dc2cb1455ae450502_main_3runs.zip: 0a6f08d075be72223edadaa4876139335183c5d23a542a9390fa053f285bbbf003_inference_routing.zip: 4a80127d10889cf70d9167f8bad60ddddcc536764e7b4d4d4ffbab54aa8da05b04_jetson_full98.zip: 4fc3b1e66734a3ac65f18c06b5ef5816610d4d676edbe01f0db11f6d3b74f5bb Research data and derived tables in these archives are licensed CC BY 4.0. Associated code and model-asset documentation are available at https://github.com/YuyangBaoo/InfraSegFM under the repository’s applicable licenses. Dataset providers’ terms apply to their source images and annotations. The all-versions DOI is 10.5281/zenodo.21510680.