3DSRBench
收藏资源简介:
3DSRBench是一个全面的3D空间推理基准数据集,由约翰斯·霍普金斯大学等机构创建。该数据集包含2,772个手动标注的视觉问答对,涵盖12种问题类型,涉及高度、位置、方向和多对象推理等4个主要类别。数据集的内容来自MSCOCO和HSSD数据集,通过人工标注和多视角合成图像生成。创建过程包括平衡数据分布、避免简单答案问题以及采用FlipEval策略进行评估。该数据集主要用于评估和提升大模态模型(LMMs)在3D空间推理任务中的表现,特别是在自动驾驶、机器人和AR/VR等领域的应用。
3DSRBench is a comprehensive 3D spatial reasoning benchmark dataset created by institutions including Johns Hopkins University. It contains 2,772 manually annotated visual question-answer pairs, covering 12 question types spanning four core categories: height, position, orientation, and multi-object reasoning. The dataset content is sourced from the MSCOCO and HSSD datasets, and is generated through manual annotation and multi-view synthetic image generation. Its development process includes balancing the data distribution, eliminating questions with overly simple answers, and adopting the FlipEval strategy for evaluation. This dataset is primarily used to evaluate and enhance the performance of large multimodal models (LMMs) on 3D spatial reasoning tasks, especially for applications in autonomous driving, robotics, and AR/VR domains.

- 13DSRBench: A Comprehensive 3D Spatial Reasoning Benchmark约翰斯·霍普金斯大学, 卡内基梅隆大学, DEVCOM陆军研究实验室 · 2024年



