UrbanVideo-Bench
收藏资源简介:
UrbanVideo-Bench是一个针对城市3D开放空间中运动认知的基准数据集。该数据集由清华大学研究团队创建,包含1500个视频片段,这些视频片段通过人工操控无人机在真实城市和模拟环境中收集得到。数据集旨在评估大型多模态模型在连续第一人称视觉观察下的运动认知能力,包括回忆、感知、推理和导航等方面。数据集覆盖了复杂城市场景和丰富的语义信息,通过设计不同类型的任务,如轨迹描述、序列回忆、目标检测、导航规划等,全面评估模型在理解城市空间运动方面的智能水平。
UrbanVideo-Bench is a benchmark dataset for motion cognition in urban 3D open spaces. Developed by a research team from Tsinghua University, it contains 1500 video clips collected via manually piloted drones in both real urban and simulated environments. This dataset aims to evaluate the motion cognition capabilities of large multimodal models during continuous first-person visual observations, covering aspects such as recall, perception, reasoning, and navigation. Covering complex urban scenarios and rich semantic information, UrbanVideo-Bench comprehensively assesses the intelligent level of models in understanding urban spatial motion by designing various types of tasks including trajectory description, sequence recall, object detection, navigation planning, and so on.

- 1UrbanVideo-Bench: Benchmarking Vision-Language Models on Embodied Intelligence with Video Data in Urban Spaces清华大学 · 2025年



