Independent Category Classifiers for Emergency Scene Description using Deep Learning approaches.

In this work, we present our proposed model for Disaster Scene Description and Indexing (DSDI) Challenge of TRECVIDI2020. For the challenge we used: the LADI Dataset, a dataset composed of scrapped Google Images using as a keyword the name of the label and an extended set of the LADI dataset. This extension was created using a crowdsourcing service like Amazon Mechanical Turk. As approaches we tried different combination of Convolutional Neural Networks (CNN), what worked better for us was using five different classifiers, one for each category of the LADI dataset. We used this configuration because we noticed that dividing the task lead to better scores. Indeed by checking the results it is possible to notice that dividing the task help the model to learn specific features for that category. We found that the dataset is very challenging and it is difficult for a model trained end to end to learn all the features useful to detect a class. For this reason, an ensemble model approach worked better for the challenge. We think that more sophisticated label for example segmentation map could have allowed obtaining better results.