Abstract
High-resolution and variable-shape images have not yet been properlyaddressed by the AI community. The approach of down-sampling data often usedwith convolutional neural networks is sub-optimal for many tasks, and has toomany drawbacks to be considered a sustainable alternative. In sight of theincreasing importance of problems that can benefit from exploitinghigh-resolution (HR) and variable-shape, and with the goal of promotingresearch in that direction, we introduce a new family of datasets (MetH). Thefour proposed problems include two image classification, one image regressionand one super resolution task. Each of these datasets contains thousands of artpieces captured by HR and variable-shape images, labeled by experts at theMetropolitan Museum of Art. We perform an analysis, which shows how theproposed tasks go well beyond current public alternatives in both pixel sizeand aspect ratio variance. At the same time, the performance obtained bypopular architectures on these tasks shows that there is ample room forimprovement. To wrap up the relevance of the contribution we review the fields,both in AI and high-performance computing, that could benefit from the proposedchallenges.