SURPRISE3D
收藏资源简介:
SURPRISE3D是一个用于评估复杂3D场景中语言引导的空间推理分割的数据集。该数据集由来自ScanNet++ v2的900多个详细室内场景中的超过20万个视觉语言对组成,包括超过2800个独特的物体类别。数据集包含8.9万多个人类注释的空间查询,这些查询在空间理解中避免了对象名称,从而减轻了捷径偏差。这些查询全面涵盖了各种空间推理技能,如相对位置、叙事视角、参数化视角和绝对距离推理。SURPRISE3D旨在促进空间感知人工智能的进步,为有效的具身交互和机器人规划铺平道路。
SURPRISE3D is a dataset for evaluating language-guided spatial reasoning segmentation in complex 3D scenes. It consists of over 200,000 visual-language pairs sourced from more than 900 detailed indoor scenes in ScanNet++ v2, encompassing over 2,800 unique object categories. The dataset contains more than 89,000 human-annotated spatial queries that avoid using object names during spatial understanding, thereby mitigating shortcut bias. These queries comprehensively cover diverse spatial reasoning skills, such as relative position, narrative perspective, parametric perspective, and absolute distance reasoning. SURPRISE3D aims to advance the progress of spatially aware AI, paving the way for effective embodied interaction and robotic planning.
Surprise3D数据集概述
数据集简介
- 名称: Surprise3D: A Dataset for Spatial Understanding and Reasoning in Complex 3D Scenes
- 作者: Jiaxin Huang, Ziwen Li, Hanlue Zhang
- 机构: MBZUAI
- 论文链接: https://arxiv.org/abs/your_arxiv_id
数据集特点
- 设计目的: 评估复杂3D场景中语言引导的空间推理分割能力
- 核心优势: 通过排除对象名称的查询设计,强调空间推理,减少语义线索带来的偏差
- 数据规模:
- 200k+ 视觉-语言对
- 900+ 室内场景(来自ScanNet++)
- 89k+ 人工标注的空间查询
- 2.8k 独特对象类别
空间推理技能覆盖
- 相对位置(如"Find the object behind the chair.")
- 叙述视角(如"Locate the object visible from the sofa.")
- 参数化视角(如"Select the object 2 meters to the left of the table.")
- 绝对距离推理(如"Identify the object exactly 3 meters in front of you.")
数据分析
- 针对低频对象的增强
- 按问题类型划分的对象频率分析(Top 15对象)
- 问题类型分布可视化
数据下载
- 下载地址: https://huggingface.co/datasets/hhllzz/surprise-3d
训练与评估
- 基于Reason3D代码库修改
- 支持ScanNet++数据预处理
- 支持Reason3D在Surprise3D上的分割任务
- 相关脚本路径: ./Models/reason3d
致谢
- 感谢Reason3D(https://github.com/KuanchihHuang/Reason3D)的作者提供了基础工作




