TTPLA: An Aerial-Image Dataset for Detection and Segmentation of Transmission Towers and Power Lines

1 Introduction

• UAV-based power-grid inspection needs precise, real-time scene understanding for localization, recognition, tracking, monitoring, inspection, and flight safety.UAV inspection is described as lower-cost, more mobile, and more flexible than traditional inspection methods.Source: §1 · p. 1
• Detection is complicated by background-color similarity, tower–line scale and class imbalance, crowded or overlapping lines, and varied object structures and shapes.Each tower is linked to at least 3–4 and up to 10 power lines; power lines are long and thin, while towers vary in shape and line-structure sparsity.Source: §1 · p. 2
• Existing TT/PL datasets support image-level or semantic annotations, while instance segmentation can distinguish same-class objects and provide masks for individual instances.The paper identifies close, occluded, and overlapping power lines as a setting where instance-level distinctions are useful.Source: §1 · p. 2
• 22.96% for bounding boxes and 15.72% for masks are the best average scores reported for the Yolact baseline, and are low in general.The baseline was evaluated with ResNet-50 and ResNet-101 backbones at three image resolutions.Source: §1 · p. 2
• TTPLA contributes a public aerial-image dataset, novel collection and labeling policies, pixel-wise instance annotations, and a baseline evaluation.The dataset includes TTs and PLs across different views, times, locations, backgrounds, and tower structures; annotations are provided in COCO format.Source: §1 · p. 3

2 Related Work

• The review describes a continuing lack of TT/PL training datasets, citing the difficulty of collecting images, especially when UAVs fly close to power grids.It surveys datasets on towers, power lines, and insulators.Source: §2 · p. 3

2.1 Datasets Based on TTs Images Only

• TT-only datasets include reduced-resolution aerial images with bounding-box labels and damage datasets, while several reviewed datasets are not publicly available.One dataset contains 3,200 images at 64×128 pixels; another has 28,674 images annotated with bounding boxes at 6,048×4,032 pixels.Source: §2.1 · p. 4

2.2 Datasets Based on PLs Images Only

• PL-only datasets include image-level or pixel-level labels, cropped UAV scenes, and synthetic wires; the reviewed synthetic datasets are not public.Manual cropping can remove realistic backgrounds and may not reflect real-time UAV operations, where objects can appear anywhere in an image.Source: §2.2 · pp. 4–5

2.3 Datasets Based on Both TTs and PLs Images

• ImageNet includes TT/PL labels, but most images are not aerial and lack top and side views, with simpler backgrounds than the UAV setting considered.The review notes 1,290 annotated images with TT and PL labels.Source: §2.3 · p. 5

3.1 Aerial Videos Collection and Images Preparation

• TTPLA draws from 80 UAV videos recorded at 3,840 × 2,160 and samples frames before manual inspection to remove redundancy and retain clear, complete tower views.Frames are sampled once every 15 frames, with more frequent sampling available for augmentation.Source: §3.1 · p. 6, §3.2 · p. 8

3.2 TTPLA Dataset Policy

• Recording policies vary tower viewpoints, zoom, time, and weather, while including noisy backgrounds, similar line-background colors, and lines partly hidden by trees.Front, top, and side views are included to support tower detection across viewing angles.Source: §3.2 · p. 6, Figure 3 · p. 7
• Expert annotators use polygons to label each tower and power-line instance at pixel level, enabling instances to be distinguished rather than assigning one class label to all matching pixels.Three expert annotators label images, with each image assigned to one annotator.Source: §3.2 · p. 8
• Tower labels group structures by lattice type or pole material rather than shape, since partial views can make tower shapes hard to distinguish; difficult-to-recognize instances are marked void and excluded from evaluation.The labels include tower-lattice, tower-tucohy, and tower-wooden, alongside cable for power lines.Source: §3.2 · pp. 8–9

3.3 Dataset Statistics

• Power-line instances greatly outnumber tower instances, yet towers and power lines occupy comparable pixel area; 659 images contain 1–6 instances and 241 contain 11–56.The authors attribute the instance imbalance to towers connecting to 2–10 power lines and include images with multiple towers to increase tower instances.Source: §3.3 · p. 10

4.1 Metrics

• Average precision is evaluated for bounding boxes and instance masks, using IoU thresholds above 50% and averages across thresholds from 50% to 95%.AP50 and AP75 use the respective 50% and 75% overlap values.Source: §4.1 · pp. 10–11
• Fig. 6 plots images against instances per image for all instances and for cable, tower-lattice, tower-tucohy, and tower-wooden classes.Source: Figure 6 · p. 11

4.2 Loss Function

• The baseline training objective sums localization, classification, and mask losses, weighted by 1, 1.5, and 6.125, with ground-truth box area normalizing mask loss.The number of boxes matching ground truth is also used in the loss formulation.Source: §4.2 · p. 11
• Classification uses softmax confidence scores, localization uses Smooth L1 box-offset loss, and mask prediction uses pixel-level binary cross-entropy.Source: §4.2 · pp. 11–12

4.3 Baseline experiment results

• Up to 1.3% of test-set images show classification falseness, including confusion between small tower-tucohy towers and PLs.The authors suggest color and shape similarity as one possible explanation.Source: §4.3 · pp. 12–13, Figure 7 · p. 12
• Segmentation falseness includes incomplete masks and split detections; PLs only 1–3 pixels wide are especially difficult for Yolact to cover fully.The account also notes that a PL may be represented by one bounding box and mask that misses much of its length.Source: §4.3 · p. 13, Figure 9 · p. 13
• 48.9%, 36.8%, 17.9%, and 2.5% of bounding-box pairs overlap at thresholds 0.3, 0.5, 0.75, and 0.95, respectively; NMS trades false positives against missed true objects.The overlap figures concern bounding boxes of different instances in TTPLA.Source: §4.3 · p. 14, Table 4 · p. 14
• The analysis highlights challenges in processing real-time UAV images and identifies opportunities for dataset-related improvement.Source: §4.3 · p. 14

• TTPLA contains 1,100 aerial images at 3,840×2,160 resolution and up to 8,987 instances, with pixel-wise annotations for instance segmentation.The dataset focuses on combined transmission towers and power lines.Source: §5 · p. 14
• The authors propose collection, annotation, and labeling policies to capture variation in view angles, scales, backgrounds, lighting, and zoom levels.Source: §5 · p. 14
• The paper establishes a baseline on TTPLA using a state-of-the-art learning model, different backbones, and varied image sizes.The authors present TTPLA as a challenge for computer vision research in detection, classification, and instance segmentation.Source: §5 · p. 14