MoF-Image: Generating Mixture-of-Features Video Game Image Dataset via GPU Rendering Simulation

Yu Wen, Xingke Yang, Aamir Bader Shah, Ruizhi Cao, Miao Pan, Chenhao Xie, Xin Fu; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2025, pp. 586-593

The video game industry is one of the largest sectors in entertainment, and computer vision (CV) has become an essential tool for enhancing game graphics and developing interactive content. Despite the growing demand for high-quality datasets to train and evaluate CV models, publicly available datasets that meet the diverse requirements of video game-related tasks remain limited. Traditional methods often focus on a single application, such as object detection or segmentation, and require manual annotation, making them unsuitable for broader CV applications. In this paper, we present a novel approach to dataset generation using a GPU simulator, which directly captures draw calls from the rendering pipeline, enabling customizable extraction of key information without relying on traditional rendering engines. Our method provides multi-resolution frames along with essential rendering details, including depth, normal direction, and object edges. We introduce MoF-Image, a new dataset generated using this approach, designed for a variety of CV tasks such as super-resolution, 3D reconstruction, and object detection. Finally, we discuss the potential applications of the MoF-Image dataset, analyze the advantages and limitations of our method, and outline directions for future improvements. The dataset is available at the following link: https://www.kaggle.com/datasets/19684b7cee0ea0e51589d1a064446c2ac72e5167a3da9732f082463e2da84821/