Use of data augmentation with Generative Adversarial Networks (GANS) to improve object detection

Abstract

Currently, object detection is performed most often using Deep Learning models. Although this type of model performs well in different scenarios, one of its problems is the need for a large and representative dataset for training. In cases where the domain of the data seen while using the detector is slightly different from the data present in training, its performance may drop. In these cases, it may be necessary that the data used for training also present images of the second domain, although its object is the same in both(detection of the same objects). Considering that the acquisition and annotation of new data is an expensive and time-consuming process, this work evaluated the use of an adversarial generative model to perform the translation of data from one domain to another, thus performing data augmentation. With the augmented data, the test was performed on an object detector trained with and without the new data, and its performance was compared. The results were promising, showing an improvement in the accuracy of the detector on large objects, but also a worse performance on small objects, given the difficulty encountered by the translation model when translating small objects from one domain to another.