Robust CNNs for detecting collapsed buildings with crowd-sourced data

Wildfires are increasingly common and responsible for widespread property damage and loss of life. Rapid and accurate identification of damage to buildings and other infrastructure can heavily affect the efficacy of disaster response during and after a wildfire. We have developed a dataset and a convolutional neural network-based object detection model for rapid identification of collapsed buildings from aerial imagery. We show that a baseline model built with crowd-sourced data can achieve better-than-chance mean average precision of 0.642, which can be further improved to 0.733 by constructing a new, more robust loss function.