UAV Bridge Defect Image Dataset and Evaluation Outputs for Keypoint-Guided Image Selection

This dataset supports Experiment 3, “Effectiveness of the Keypoint-Guided Image Selection,” in a study of UAV-based bridge inspection using vision–language models. The dataset contains 1,581 original UAV bridge images, 79 images retained after keypoint-guided image selection, expert ground-truth annotations for Crack, Spalling, and Corrosion, frozen Qwen VLM outputs for the original and selected image sets, and geometric clustering summaries associated with the selected image set. A VLM prediction is considered correct when its predicted defect class matches the ground-truth class and its normalized center point lies inside the corresponding ground-truth bounding box. Multiple same-class prediction points may match the same ground-truth box. The source code used for Qwen-based defect analysis and instance-level evaluation is archived separately. The source code of the keypoint-guided image-selection module is not included. The final selected image subset, clustering summaries, frozen VLM outputs, and evaluation code are provided.