Abstract
The ability to classify images is dependent on having access to large labeleddatasets and testing on data from the same domain that the model can train on.Classification becomes more challenging when dealing with new data from adifferent domain, where gathering and especially labeling a larger imagedataset for retraining a classification model requires a labor-intensive humaneffort. Cross-domain classification frameworks were developed to handle thisdata domain shift problem by utilizing unsupervised image-to-image translationmodels to translate an input image from the unlabeled domain to the labeleddomain. The problem with these unsupervised models lies in their unsupervisednature. For lack of annotations, it is not possible to use the traditionalsupervised metrics to evaluate these translation models to pick the best-savedcheckpoint model. This paper introduces a new method called Domain-knowledgeInspired Pseudo Supervision (DIPS) which utilizes domain-informed GaussianMixture Models to generate pseudo annotations to enable the use of traditionalsupervised metrics. This method was designed specifically to supportcross-domain classification applications contrary to other typically usedmetrics such as the FID which were designed to evaluate the model in terms ofthe quality of the generated image from a human-eye perspective. DIPS provesits effectiveness by outperforming various GAN evaluation metrics, includingFID, when selecting the optimal saved checkpoint model. It is also evaluatedagainst truly supervised metrics. Furthermore, DIPS showcases its robustnessand interpretability by demonstrating a strong correlation with trulysupervised metrics, highlighting its superiority over existing state-of-the-artalternatives. The code and data to replicate the results can be found on theofficial Github repository: https://github.com/Hindawi91/DIPS