PointArena
收藏资源简介:
PointArena是一个全面的多模态指点点评价平台,由华盛顿大学、艾伦人工智能研究所和安德森学院职业技术学院的研究团队开发。该平台包含三个部分:Point-Bench,一个包含约1000个指点点任务的定制数据集,分为五个推理类别;Point-Battle,一个交互式、基于网络的竞技场,用于模型之间的匿名、成对比较,已经收集了超过4500个匿名投票;Point-Act,一个现实世界的机器人操作系统,允许用户直接评估多模态模型在现实场景中的指点点能力。PointArena旨在通过语言指导的多模态指点点任务,评估多模态模型的空间定位精度,并支持下游应用,如机器人技术、增强现实和人机交互。
PointArena is a comprehensive multimodal pointing evaluation platform developed by research teams from the University of Washington, the Allen Institute for AI, and Anderson College of Career and Technical Education. The platform consists of three parts: Point-Bench, a custom dataset containing approximately 1,000 pointing tasks divided into five reasoning categories; Point-Battle, an interactive web-based arena for anonymous pairwise comparisons between models, which has collected over 4,500 anonymous votes; and Point-Act, a real-world robotic operating system that allows users to directly evaluate the pointing capabilities of multimodal models in realistic scenarios. PointArena aims to evaluate the spatial localization accuracy of multimodal models via language-guided multimodal pointing tasks, and supports downstream applications such as robotics, augmented reality, and human-computer interaction.
Point Arena 数据集概述
数据集简介
- 名称:Point Arena
- 研究主题:通过语言引导的指向任务探究多模态基础能力
- 核心目标:评估多模态大语言模型(MLLMs)中语言与视觉的精确空间对齐能力
- 特点:
- 首个专门评估语言引导指向能力的开放统一平台
- 提供标准化场景、多样化数据集和严格评估协议
- 填补现有基准测试在细粒度基础任务上的空白
数据集组成
1. Point-Bench
- 功能:语言与视觉间精确空间对齐的标准化评估
- 评估维度:
- Affordance
- Spatial
- Reasoning
- Steerability
- Counting
- Average
- 排名示例:
- 第1名:Human (平均分89.128)
- 第2名:Molmo-72B (平均分63.832)
- 第3名:Molmo-7B-O (平均分63.266)
2. Point-Battle
- 功能:评估不同模型类型和提示策略的性能差异
- 评估指标:
- Elo Rating
- Wins
- Losses
- Games
- Win Rate
- Confidence Interval
- 排名示例:
- 第1名:allenai/Molmo-7B-D-0924 (Elo 1205.3)
- 第2名:Qwen/Qwen2.5-VL-7B-Instruct (Elo 1058.9)
3. Point-Act
- 功能:提供标准化场景和严格评估协议的多样化数据集
- 评估指标:
- Success Rate
- SUS Score
- 示例结果:
- Human: 成功率90%, SUS评分88.2
- Molmo: 成功率70%, SUS评分61.6
数据获取
- Point-Bench数据:下载完整数据(CSV)
- Point-Battle数据:下载完整数据(CSV)
- Point-Act数据:下载完整数据(CSV)
研究团队
- 机构:
- 华盛顿大学
- 艾伦人工智能研究所
- Anderson Collegiate Vocational Institute
- 主要作者:
- Long Cheng (共同第一作者)
- Jiafei Duan (共同第一作者)
- Yi Ru Wang (共同第二作者)
- Haoquan Fang (共同第二作者)
- Boyang Li (共同第二作者)
引用信息
bibtex @misc{cheng2025pointarenaprobingmultimodalgrounding, title={PointArena: Probing Multimodal Grounding Through Language-Guided Pointing}, author={Long Cheng and Jiafei Duan and Yi Ru Wang and Haoquan Fang and Boyang Li and Yushan Huang and Elvis Wang and Ainaz Eftekhar and Jason Lee and Wentao Yuan and Rose Hendrix and Noah A. Smith and Fei Xia and Dieter Fox and Ranjay Krishna}, year={2025}, eprint={2505.09990}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2505.09990}, }



