SuperMarioDomains: Generalizing to Domains with Evolving Graphics

Abstract

Domains in previous Domain Generalization (DG) benchmarks have been sampled from various image collections of different styles such as photographs, sketches, cartoons, paintings, product images, and etc. However, from these existing DG datasets, it is still difficult to quantify the magnitude of domain shift between different domains and relate that to the performance gap across domains. It is also unclear how to measure the overlap between different domains. Therefore, we present a new DG dataset, SuperMarioDomains, containing four domains that are derived from four chronological titles in the Mario video game franchise on four generations of video game hardware. The discrepancy between our domains is quantified in terms of image representation complexity that reflect the hardware evolution in image resolution, color palette, and presence of 3D rendering. We benchmark state-of-the-art DG algorithms under both Multi-Source and Single-Source DG settings on our dataset and find that they can only surpass the random average baseline in our dataset by at most 18.0% and 10.4% respectively. In addition, we show that adding our dataset as part of the pre-training process improves performance of existing DG algorithms on the PACS benchmark.