DETR-based model detects utility poles with 90.43% mAP and estimates lean angle

Automating utility pole inspection just got more precise. Researchers Carl Dickinson and Gaetano Di Caterina, publishing at BMVC 2025, present a deep learning framework built on the Detection Transformer (DETR) architecture that detects wooden utility poles, segments them, estimates their lean angle, and classifies attached electrical warning signs from ground-level images. The model was trained on a custom dataset of 4,570 annotated images drawn from Google Street View, deliberately featuring visually ambiguous poles that lack distinctive features—a challenge for traditional detectors.

The results beat standard object detectors like RetinaNet, Faster R-CNN, and YOLOv3-Tiny. The DETR-based model achieves 90.43% mean average precision (mAP) for pole detection and 88.26% mAP for sign detection. By adding a segmentation head, the system generates per-instance masks, which are used to estimate pole lean angle. It accurately estimated lean for 1,367 out of 1,433 test-set poles, with a mean absolute error of just 1.01 degrees. The dataset and model are publicly available, providing a benchmark for future infrastructure-inspection research—and a practical tool for grid monitoring.