On the Domain Generalization Capabilities of Interactive Segmentation Methods

Interactive image segmentation (IIS) methods are usually trained over segmentation datasets containing natural images. They are also usually evaluated over natural images. However, the most common use case is the annotation of new images from a different domain. Yet, the performance of IIS methods on a different domain is seldom reported. In this work, we evaluate a state-of-the-art IIS method trained with natural images over an aerial image dataset. Its performance is compared to the performances the method achieves when being trained/finetuned with aerial images. The comparison reveals that there is a big domain generalization gap.