Wireless Leiden Site Photos Orientation Dataset

A privacy-preserving dataset of technical site photographs documenting wireless network infrastructure — antennas, masts, node installations, site surveys, building projects, and equipment deployments. Each image is labeled with a human-verified orientation label (0, 90, 180, or 270 degrees) indicating the rotation needed to display the image correctly. All images have been screened for the presence of recognizable persons using YOLOv8 object detection and manually verified; no images containing persons are included. The dataset includes CLIP-based orientation predictions with confidence scores for comparison. Images retain their EXIF metadata, and the directory structure (site/collection) preserves the original temporal and spatial grouping of the archive. Custom labeling tools are included. Evaluation: effect of orientation correction on privacy detection. YOLOv8 person detection was compared across three image versions: original (with EXIF orientation), human-corrected rotation, and CLIP-predicted rotation. Human-corrected rotation significantly improved YOLOv8 recall from 0.673 to 0.927 (+0.255) and reduced false negatives by 78% (54 to 12), confirming that upright persons are far easier to detect than sideways or upside-down ones. In contrast, automated CLIP rotation without human correction was harmful: false negatives tripled (88 to 276), recall dropped from 0.836 to 0.487, and F1 fell by 0.25, because CLIP predicted the wrong rotation in 98.1% of rotated images. The recommended pipeline is: CLIP pre-screening, human review, then YOLOv8 privacy filtering. Evaluation scripts and raw results are included in the dataset.