VSI-Bench
收藏资源简介:
VSI-Bench是一个用于评估多模态大语言模型(MLLMs)视觉空间智能的新型视频基准数据集,由纽约大学、耶鲁大学和斯坦福大学联合创建。该数据集包含超过5000个问题-答案对,来源于288个真实的室内场景视频,涵盖住宅、办公室和工厂等多种环境。数据集的创建过程结合了现有的3D重建和理解数据集,如ScanNet、ScanNet++和ARKitScenes,并通过人工审核确保数据质量。VSI-Bench主要用于测试模型在空间感知、物体识别和路径规划等任务中的表现,旨在推动多模态模型在视觉空间智能领域的应用和发展。
VSI-Bench is a novel video benchmark dataset for evaluating the visual spatial intelligence of multimodal large language models (MLLMs), jointly developed by New York University, Yale University, and Stanford University. This dataset contains over 5,000 question-answer pairs, sourced from 288 real indoor scene videos covering diverse environments including residential spaces, offices, and factories. The construction of VSI-Bench integrates existing 3D reconstruction and understanding datasets such as ScanNet, ScanNet++, and ARKitScenes, and ensures data quality through manual review. VSI-Bench is primarily used to test model performance on tasks including spatial perception, object recognition, and path planning, aiming to promote the application and development of multimodal models in the field of visual spatial intelligence.

- 1Thinking in Space: How Multimodal Large Language Models See, Remember, and Recall Spaces纽约大学、耶鲁大学、斯坦福大学 · 2024年



