D-ALICE: Domain Adaptation-based Labeling the human In Cartoon imagE

How did Alice, who went to the Wonderland, solve the problems? In this paper, as Alice solved problems by changing the size of her body in the Wonderland, we classified the person by changing the style using the image translation technique for the cartoon image in pretrained segmentation model. In general, when you test a cartoon image on a pretrained segmentation model based on real image, the results do not appear correctly. To solve this problem, the ground truth for new images should be created and newly trained. This approach is costly and time consuming. So, we propose a method based on domain adaptation to label the human in cartoon image (D-ALICE) without training a new segmentation model by transforming images using a CycleGAN-based model that can be trained with an unpaired dataset. The quantitative and qualitative evaluation of pre and post conversion images resulted from the segmentation model trained as MIT ADE20K were conducted, and the mean-IoU was increased by more than 35%. The results of this research can be applied to other domains without newly training the deep learning model, and furthermore it can help to provide the ground truth for the data which does not have ground truth which does not have before.