Abstract
Images or video recordings assist emergency responders in quickly inspecting the damage after a disaster event. New techniques are needed to help responders organize and find important information at the right time. However, most existing methods do not meet public safety standards due to a lack of training data. We propose a multi-source weak supervision fusion technique to train on a highly imbalanced dataset annotated with noisy labels. Using a Confident Learning technique, we reduce the noise effect while boosting the class labels' quality. We combine the predictive power from models trained on large-scale visual datasets using Differential Evolution. This research demonstrates a fully-automatic approach with great potential to reduce the required time and resources while delivering exceptional results. In the TRECVID2021 Disaster Scene Description and Indexing (DSDI) Challenge, our technique achieved the top score among all the submitted runs, independent of the training data utilized.