SpatialUAV
收藏资源简介:
SpatialUAV是一个用于评估真实低空无人机(UAV)场景中空间智能的诊断性基准数据集,专注于视觉问答任务,旨在测试模型在复杂无人机感知环境下的空间推理能力。数据集共包含4,331个精心构建的视觉问答实例,覆盖14种不同的任务类型,主要分为五大类:语义判别(例如区域识别)、空间关系理解、无人机-无人机协作(空-空协作)、无人机-地面协作(空-地协作)以及运动理解。它采用统一的“视觉输入-问题-答案”模式,提供七种不同的视觉输入配置(例如单张图像、多视角图像序列、视频帧序列)和九种答案格式,数据规模属于1K到10K的范畴。数据集仅包含一个test拆分,专门用于基准测试和模型评估,并提供两种配置:full配置包含全部4,331个实例,用于完整评估;subset_20pct_per_task配置从每种任务类型中抽取20%的样本,共863个实例,用于快速开发和建立人工参考基准。每个样本由五个字段构成:唯一标识符id(其前缀指明了任务类型)、有序的图像或视频帧路径列表image、采用{from, value}对话格式存储的问题或指令conversations、标识数据集来源的标签source(固定为“SpatialUAV”)以及标准的真实答案GT。视觉数据(图像/视频帧)需要从压缩包中单独下载并解压,解压后包含多个子目录,对应不同的任务和视觉输入配置。该数据集整合了来自BEDI、AirCopBench、MAVREC、AirScape、University-1652等多个上游数据集的数据,因此使用需遵守这些原始数据集的许可证和条款。
SpatialUAV is a diagnostic benchmark dataset for evaluating spatial intelligence in real-world low-altitude unmanned aerial vehicle (UAV) scenarios. It focuses on visual question answering tasks, aiming to test models spatial reasoning capabilities in complex UAV perception environments. The dataset contains a total of 4,331 carefully constructed visual question answering instances, covering 14 different task types. These task types are mainly divided into five categories: semantic discrimination (e.g., region recognition), spatial relationship understanding, UAV-UAV collaboration (air-air collaboration), UAV-ground collaboration (air-ground collaboration), and motion understanding. The dataset designs a unified visual input-question-answer pattern, providing seven different visual input configurations (e.g., single image, multi-view image sequences, video frame sequences) and nine answer formats. The data scale falls within the 1K to 10K range. The dataset includes only a `test` split, specifically for benchmark testing and model evaluation. For convenience, two configurations are provided: the `full` configuration contains all 4,331 instances for complete benchmark evaluation, while the `subset_20pct_per_task` configuration samples 20% from each task type, totaling 863 instances, for rapid development and establishing human reference benchmarks. Each data sample consists of five fields: a unique identifier `id` (whose prefix indicates the task type), an ordered list of image or video frame paths `image`, questions or instructions stored in a `{from, value}` conversation format `conversations`, a label `source` identifying the dataset source (fixed as SpatialUAV), and the standard ground truth answer `GT`. Visual data (images/video frames) need to be downloaded separately from a compressed package and extracted, containing multiple subdirectories corresponding to different tasks and visual input configurations. The dataset integrates data from multiple upstream datasets such as BEDI, AirCopBench, MAVREC, AirScape, and University-1652, so its use must comply with the licenses and terms of these original datasets.
数据集名称
SpatialUAV
数据集简介
SpatialUAV 是一个用于评估低空无人机(UAV)场景中空间智能的诊断性基准数据集。它包含 4,331 个精心策划的视觉问答实例,覆盖 14 种任务类型,涉及语义判别、空间关系、空中-空中协作、空中-地面协作和运动理解。
语言与任务
- 语言:英语
- 任务类别:视觉问答(Visual Question Answering)
- 标签:无人机、空间推理、多视图、空中-地面、运动理解、基准测试
数据集规模
- 总实例数:4,331
- 数据大小范围:1K < 样本数 < 10K
数据集配置
| 配置名称 | 分割 | 行数 | 用途 |
|---|---|---|---|
full |
test |
4,331 | 完整的基准评估 |
subset_20pct_per_task |
test |
863 | 更快的开发和人工参考子集 |
数据记录结构
每条记录包含 5 个字段:
| 字段 | 类型 | 描述 |
|---|---|---|
id |
字符串 | 唯一标识符;其前缀表示任务类型 |
image |
字符串列表 | 图像或视频帧的有序路径 |
conversations |
对象列表 | 以 {from, value} 格式存储的问题或指令 |
source |
字符串 | 数据集来源标签(SpatialUAV) |
GT |
字符串 | 标准真实答案 |
视觉输入配置
该基准提供七种视觉输入配置和九种答案格式。视觉数据需下载并解压以下文件:
samples_Single_Image/samples_A2A_Pured/samples_A2A_detected/samples_A2A_Occlusion_Removal/samples_A2G_Pured/samples_A2G_detected/samples_A2G_Path_Planning/samples_Motion_Understanding_Frames/
数据来源
数据集衍生自多个源数据集,包括 BEDI、AirCopBench、MAVREC、AirScape、University-1652 等。用户需自行审查并遵守这些源数据集的许可条款。
许可证
- 代码:MIT 许可证
- 基准数据:
license: other(未以 MIT 重新授权,受源数据集各自许可证约束)
引用
若使用 SpatialUAV,请引用以下文献:
bibtex @article{zhang2026spatialuav, title = {SpatialUAV: Benchmarking Spatial Intelligence for Low-Altitude UAV Perception, Collaboration, and Motion}, author = {Zhang, Haoyu and Liu, Meng and Xiang, Qianlong and Wang, Kun and Wang, Yaowei and Nie, Liqiang}, journal = {arXiv preprint arXiv:2606.27876}, year = {2026} }
更多信息
有关任务定义、数据集构建、评估指标、模型推理和基准结果,请访问 GitHub 仓库 和 SpatialUAV 论文。




