PulseCheck457
收藏资源简介:
PulseCheck457是一个由约翰霍普金斯大学创建的、 unbiased和可扩展的合成数据集,旨在诊断大型多模态模型(LMMs)的6D空间推理能力。该数据集包含了457个注释完整的场景,每个场景都有多种物体,并提供了关于物体的2D和3D位置、方向等详细信息。数据集针对多对象识别、2D位置、3D位置和3D方向设计了多种问题,难度分为五个级别,从单一对象识别到复杂的6D空间关系推理任务。这些问题覆盖了从基本的2D空间关系到高级的3D位置和方向理解,旨在评估模型在处理不同复杂度的空间推理任务时的性能。
PulseCheck457 is an unbiased and scalable synthetic dataset created by Johns Hopkins University, designed to diagnose the 6D spatial reasoning capabilities of large multimodal models (LMMs). This dataset contains 457 fully annotated scenes, each featuring multiple objects, and provides detailed information including the 2D and 3D positions and orientations of the objects. A variety of questions are formulated for tasks such as multi-object recognition, 2D position, 3D position and 3D orientation, with five difficulty levels ranging from single-object recognition to complex 6D spatial relationship reasoning tasks. These questions cover a broad spectrum from basic 2D spatial relationship understanding to advanced 3D position and orientation comprehension, aiming to evaluate the performance of models when handling spatial reasoning tasks of varying complexities.




