JingkunAn/RefSpatial-Expand-Bench

RefSpatial-Expand-Bench 是一个基于真实世界杂乱场景的基准测试，用于评估更复杂的、具有推理功能的多步骤空间引用。它包括两个任务：位置任务和放置任务。位置任务包含 241 个样本，需要模型预测一个 2D 点来指示唯一的靶标对象。放置任务包含 200 个样本，需要模型预测一个 2D 点在期望的自由空间内。数据集还引入了推理步骤（`step`），用于表示帮助限制搜索空间的锚点对象及其空间关系的数量。`step` 值越高，表示推理复杂性越大，需要更强的空间理解和推理能力。

RefSpatial-Expand-Bench is a benchmark for multi-step spatial referring with reasoning based on real-world cluttered scenes. It includes two tasks: Location and Placement. The Location task contains 241 samples, requiring the model to predict a 2D point indicating the unique target object. The Placement task contains 200 samples, requiring the model to predict a 2D point within the desired free space. The dataset introduces reasoning steps (`step`) for each sample, representing the number of anchor objects and their spatial relations that help constrain the search space. A higher `step` value reflects greater reasoning complexity and a stronger need for spatial understanding and reasoning.