Data for : "OHL-UK: Wooden Utility Pole and Electrical Sign Corpus with Trained Detection and Segmentation Models"

The OHL-UK dataset and trained models support the BMVC 2025 paper “Advancing Utility Pole and Sign Detection Through Deep Learning” by Carl Dickinson and Gaetano Di Caterina. The dataset consists of ground-level imagery of wooden utility poles without crossarms and their associated electrical warning signs. It was created to enable research in object detection, instance segmentation, and pole lean estimation for UK overhead line infrastructure.

Images were collected using the Google Street View API from over 670,000 geographic coordinates supplied by UK Power Networks, with four compass views captured at each point. All images are 640 × 640 pixel JPEGs and were manually filtered to remove irrelevant scenes. Two object classes are annotated: wooden poles and electrical warning signs. Annotations are provided in COCO JSON format, with the test set extended with pole lean angle information.

The dataset is organised into three components: Object Detection, Segmentation, and Test Images. A Models folder contains best-performing trained weights for DETR, DINO-DETR, Faster R-CNN, RetinaNet, YOLOv3-tiny, and YOLOv8 architectures, in formats including PyTorch, TensorFlow/Keras, and Darknet.

The dataset is openly available under a CC-BY 4.0 licence. Creator: Carl Dickinson, Department of Electronic and Electrical Engineering, University of Strathclyde.