The authors of the PLD-UAV: Power Line Detection in UAV dataset emphasize the significance of power line detection within an automated UAV-based electricity inspection system. This detection is pivotal for real-time motion planning and navigation along power lines. They address the limitations of previous methods, often reliant on traditional filters and gradients, which may falter in accurately capturing power lines amidst noisy backgrounds. To tackle this challenge, the authors have devised a precise power line detection approach leveraging convolutional and structured features. Initially, a convolutional neural network is employed to generate hierarchical responses from each layer. This method harnesses multiscale and structured prior information to execute both accurate and efficient detection.

The context of power line detection is underscored by its vital role in inspecting and maintaining power lines, crucial for ensuring their reliability and safety. Traditional inspection methods, often involving skilled linemen, are not only time-consuming but also perilous. Consequently, the development of an automated UAV-based inspection system bears the potential to save time, money, and mitigate dangerous incidents. Among the core modules of this system, power line detection stands out as pivotal, aiding in autonomous navigation by accurately identifying and locating power lines.

maxDist is a maximal tolerance for edge localization in evaluation.

The paucity of large public datasets with pixel-wise annotations has hindered the development of learning-based methods for power line detection. Previous methods often built upon traditional edge detectors and were evaluated on a limited number of images. Learning-based approaches struggled due to the lack of well-annotated datasets for training and testing. In an effort to address this issue, the authors have contributed by releasing two power line detection datasets: the Power Line Dataset of Urban Scenes (PLDU) and the Power Line Dataset of Mountain Scenes (PLDM). These datasets feature diverse sample images, as shown in Figures 4 and 5. PLDM, in particular, exhibits thinner power lines, attributed to the longer shooting distance, with backgrounds primarily comprising leaves and grasses. PLDU dataset images were captured with a UAV hovering at altitudes within ten meters above the power lines, whereas PLDM dataset images were acquired from distances exceeding thirty meters. The misalignment between manually labeled ground truth and actual boundaries is acknowledged, and a parameter called maxDist.

Data augmentation, recognized as a valuable technique for enhancing edge learning performance, was applied. The authors rotated images by 45 degrees, followed by image flips at each angle. Additionally, multiscale images, achieved by resizing training images to scales of 0.5 and 1.5, were employed during training. Consequently, the augmented training set expanded to 48 times its original size.

Notably, the distinctive aspect of these datasets lies in their real-world origin, captured via UAVs, rather than relying on synthetic data. Prior experiments have demonstrated that training solely on synthetic data falls short of achieving the performance of models fine-tuned on real-world datasets. Furthermore, the datasets encompass a diverse range of power line orientations, positions, lengths, widths, and backgrounds, making them valuable for developing learning-based methods for precise power line detection. While the PLDM dataset is relatively small for deep learning approaches, the authors primarily employ the PLDU dataset as the main dataset for evaluation.