TL;DR
This paper introduces an Adaptive Training Sample Selection (ATSS) deep learning method for automatically detecting and mapping utility poles from aerial images. ATSS outperforms existing methods like Faster R-CNN and RetinaNet, offering a more accurate and efficient solution for large-scale infrastructure mapping.
What Problem Does This Paper Address?
Maintaining and mapping utility poles is crucial for power distribution networks, but traditional methods like manual street-by-street surveying are time-consuming and costly, especially for large areas. While aerial imagery offers a more efficient alternative, detecting utility poles from top-down views is challenging due to their small size, varied appearances, and potential occlusions from trees or shadows. Existing deep learning methods, while powerful, often struggle with these challenges, leading to inaccuracies. This paper aims to bridge this gap by proposing and evaluating a novel deep learning approach, ATSS, specifically tailored for utility pole detection in aerial orthoimages.
Methods
The core of this research is the application and evaluation of the Adaptive Training Sample Selection (ATSS) deep learning method for utility pole detection. ATSS is a state-of-the-art object detection algorithm that distinguishes itself by its unique approach to selecting positive and negative training samples. Unlike traditional methods that rely on fixed Intersection over Union (IoU) thresholds, ATSS dynamically determines an individual IoU threshold for each ground-truth bounding box. This adaptive strategy is hypothesized to improve the detection of small objects or objects occupying a small portion of their bounding box, such as utility poles.
The authors compare ATSS against two widely used object detection methods in remote sensing: Faster R-CNN and RetinaNet. Both Faster R-CNN and RetinaNet are described as two-stage and one-stage detectors, respectively, utilizing a ResNet-50 backbone and Feature Pyramid Network (FPN) for feature extraction. The implementation of all methods was based on the MMDetection framework.
A key aspect of the experimental setup involved investigating the influence of bounding box size on detection performance. The authors systematically varied the bounding box dimensions from 30x30 to 70x70 pixels to understand its impact on ATSS's accuracy.
Datasets
The study utilized a dataset of aerial RGB orthoimages provided by the city hall of Campo Grande, Mato Grosso do Sul, Brazil. These orthoimages have a Ground Sample Distance (GSD) of 10 cm.
- Total Orthoimages: 1057 images with dimensions 5619 × 5946 pixels.
- Patches: The images were split into 99,473 non-overlapping patches of 256 × 256 pixels.
- Ground-truth: 111,583 utility poles were identified as ground-truth. Initially, these were labeled as points, which were then converted into bounding boxes for object detection.
The dataset was divided into training (60%), validation (20%), and testing (20%) sets.
Key Insights and Ideas
- Adaptive IoU Thresholding is Key: The ATSS method's ability to adapt the IoU threshold for each ground-truth bounding box is crucial. This allows it to better handle objects like utility poles that occupy a small fraction of their bounding box, preventing false positives from background intersections.
- Bounding Box Size Matters: The size of the bounding box used to represent the utility pole significantly impacts detection performance. Larger bounding boxes (e.g., 60x60 pixels) were found to be more effective than smaller ones (e.g., 30x30 pixels) for ATSS, likely by providing more context and including shadows that aid detection.
- Aerial Imagery is Feasible: Despite the challenges of detecting small objects in aerial views, this research demonstrates that with appropriate methods like ATSS, accurate mapping of utility poles is achievable.
- Challenging Scenarios Handled: ATSS showed superior performance in challenging scenarios like poles occluded by trees and those in shadow regions, where Faster R-CNN and RetinaNet struggled, often producing multiple bounding boxes for a single object.
Key Results
- Accuracy: ATSS achieved the highest average precision (AP) across the board.
- With a 40x40 bounding box and an IoU threshold of 0.5 (AP50), ATSS achieved 0.913, outperforming Faster R-CNN (0.875) and RetinaNet (0.874) by approximately 4%.
- With an IoU threshold of 0.75 (AP75), ATSS achieved 0.749, outperforming the others by approximately 7%.
- Bounding Box Influence: For ATSS, increasing the bounding box size from 30x30 to 60x60 pixels improved AP50 by 6.5% and AP75 by 23.1%.
- Computational Costs: All three methods exhibited similar computational costs, with an average processing time of approximately 0.048 seconds per patch, translating to an area scan speed of around 13,651 m²/s. This indicates that ATSS is not computationally prohibitive compared to existing methods.
Limitations and Future Work
- Bounding Box Size Optimization: While larger bounding boxes improved performance, very large boxes (e.g., 70x70) showed diminishing returns and could include irrelevant information. Further optimization of bounding box size or adaptive sizing could be beneficial.
- Generalizability: The study was conducted on a specific urban area in Brazil. Future work could explore the generalizability of ATSS to diverse urban environments with different pole types, infrastructure densities, and image acquisition conditions.
- Operational Tool Development: The authors suggest that ATSS is a strong candidate for developing an operational tool for automatic utility pole detection and mapping, implying potential for real-time applications or deployment on embedded systems with limited computational capacity.
Summary updated