PulseCheck457
收藏资源简介:
PulseCheck457是一个为了评估大型多模态模型在6D空间推理方面的能力而设计的合成数据集。该数据集由约翰斯·霍普金斯大学和DEVCOM陆军研究实验室创建,包含457个条目,专为诊断6D空间推理能力而设计,覆盖了多对象识别、2D位置、3D位置和3D方向四个关键空间推理能力。数据集通过合成场景实现了无偏的属性设计,并为对象提供了完整的6D注释,包括3D世界坐标和姿态方向。该数据集旨在解决多模态模型在处理复杂三维空间推理任务时的性能评估问题。
PulseCheck457 is a synthetic dataset developed to evaluate the 6D spatial reasoning capabilities of large multimodal models. It was co-created by Johns Hopkins University and the DEVCOM Army Research Laboratory. The dataset consists of 457 entries, which are purpose-built to diagnose 6D spatial reasoning abilities, covering four core spatial reasoning skills: multi-object recognition, 2D position, 3D position, and 3D orientation. It adopts synthetic scenarios to achieve unbiased attribute design, and provides complete 6D annotations for all objects, including 3D world coordinates and pose orientation. This dataset is intended to address the performance evaluation issue faced by multimodal models when handling complex 3D spatial reasoning tasks.

- 1PulseCheck457: A Diagnostic Benchmark for 6D Spatial Reasoning of Large Multimodal Models约翰斯·霍普金斯大学,DEVCOM陆军研究实验室 · 2025年



