Locating Urban Trees Near Electric Wires Using Google Street View Photos: A New Dataset and a Semi-Supervised Learning Approach in the Wild

Artur André A. M. Oliveira, Zhangyang Wang, Roberto Hirata; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2022, pp. 4286-4294

Vegetation is desirable in most urban spaces, but its management is not easy, mainly the intersection between trees and sidewalks, or trees and electric wires. This work presents a method to automatically detect the latter using ground-level images instead of aerial images. Real-world ground-level urban images are cheap to collect, but they may be hard to label and classify because neural networks tend to be overconfident, and manually labeling thousands of images may be cumbersome and unfeasible. We propose using Focal Loss to calibrate an overconfident neural network and the use of the training protocol Noisy Student to lessen the burden of manually labeling images. Our results show that these methods improve the results over the Cross-Entropy loss, and the confidence levels of the predictions can be used in an Active Learning system to improve the overall accuracy.