VLM UAV bridge images expert annotations and experimental data

This dataset accompanies the manuscript “From UAV Images to Semantically Annotated 3D Models: A Keypoint-Guided Vision–Language Model Framework for Infrastructure Inspection”. It contains 2869 original bridge photographs acquired by the authors using a DJI M4E UAV, expert defect annotations, available expert component labels, keypoint reference labels, and experimental records supporting Experiments 3–5. The release includes 1581 photographs in the original collection and 1288 photographs in the add1 collection. Experiments 3–4 use 1209 left/right photographs from original and all 1288 photographs from add1 as inputs for keypoint-guided image selection. Experiment 5 uses positive annotations from the full collection. The dataset provides 714 image-level defect bounding boxes covering crack, corrosion, and spalling. The component vocabulary contains six classes, with expert component labels supplied for the evaluated samples. Selected-view manifests, five radius configurations, saved model responses, the final 20 draws of 13 targets each, matching records, and reference metrics are included. Supplementary prompts, raw model outputs, and resource-usage records are also provided for the two heritage case studies in Experiment 2; the heritage input photographs and 3D models are not included. The companion software is available at https://doi.org/10.5281/zenodo.23051812. It provides the recorded prompts, inference settings, image-selection and inference implementations, and an offline evaluator that recomputes the reported metrics for Experiments 3–5 from the released annotations and saved model outputs without new API calls. The dataset is distributed as six independent ZIP archives. Download all six archives and extract them into the same directory, merging the SR_VLM_Data folder. Part 01 contains documentation, annotations, experimental records, checksums, and a subset of the photographs; the remaining parts contain the other photographs. All six parts are required for the complete dataset. To run the companion software, place the extracted SR_VLM_Code and SR_VLM_Data folders side by side and follow the instructions in the software README.