StarCraftImage

StarCraftImage是一个基于StarCraft II游戏回放的大型多代理空间推理基准数据集，由普渡大学创建。该数据集包含从60,000个游戏回放中提取的360万张摘要图像，每张图像都是游戏状态的详细快照，包括所有单位的位置、单位ID以及重要元数据如游戏结果和玩家种族。数据集设计了三种不同复杂度的图像表示：高光谱图像、RGB图像和灰度图像，旨在简化空间推理任务的原型设计和方法测试，同时保持复杂和战略性的对象定位。数据集适用于多种应用，如自主监控传感器网络和强化学习子任务，旨在解决复杂空间推理问题。

StarCraftImage is a large-scale multi-agent spatial reasoning benchmark dataset based on StarCraft II game replays, developed by Purdue University. This dataset contains 3.6 million summary images extracted from 60,000 game replays, where each image serves as a detailed snapshot of the game state, including the positions of all in-game units, unit IDs, and critical metadata such as game outcomes and player races. Three image representations with varying complexity levels are offered in the dataset: hyperspectral images, RGB images, and grayscale images. It is intended to simplify the prototyping and method testing of spatial reasoning tasks while preserving the complex and strategic nature of object localization. The dataset supports a wide range of applications such as autonomous monitoring sensor networks and reinforcement learning subtasks, and aims to solve complex spatial reasoning problems.