Paper2Poster
收藏资源简介:
Paper2Poster数据集由滑铁卢大学、新加坡国立大学、牛津大学的研究团队创建,旨在评估和推进学术海报自动生成技术。该数据集包含了100篇AI领域的会议论文及其对应的作者设计的海报,数据量丰富,涵盖了计算机视觉、自然语言处理、强化学习等多个子领域。Paper2Poster数据集的创建过程严格遵循了科学研究的规范,保证了数据的高质量和多样性。该数据集主要用于解决学术海报自动生成中的挑战,例如长文本上下文压缩、多模态输入的处理以及布局感知的多模态输出等。
The Paper2Poster dataset was developed by research teams from the University of Waterloo, National University of Singapore, and University of Oxford, with the aim of evaluating and advancing automatic academic poster generation technologies. This dataset includes 100 AI conference papers and their corresponding posters designed by the original authors, featuring abundant data volume and covering multiple subfields such as computer vision, natural language processing, and reinforcement learning. The creation of the Paper2Poster dataset strictly adheres to scientific research norms, ensuring the high quality and diversity of the dataset. This dataset is primarily employed to address key challenges in automatic academic poster generation, such as long-text context compression, multimodal input processing, and layout-aware multimodal output.
Paper2Poster: Multimodal Poster Automation from Scientific Papers
数据集概述
- 名称: Paper2Poster
- 类型: 多模态海报生成与评估数据集
- 主要功能: 从科学论文自动生成海报并评估海报质量
核心组件
-
PosterAgent: 自上而下、视觉在环的多智能体系统,实现从
paper.pdf到poster.pptx的转换- 包含Parser、Planner、Painter-Commentor三个模块
-
评估体系:
- PaperQuiz: 通过问答测试海报传达论文核心内容的能力
- VLM-as-Judge: 使用视觉语言模型评估
- 统计指标: 视觉相似度、PPL等
技术特性
- 支持多种模型组合:
- GPT-4o
- Qwen-2.5-7B-Instruct
- 其他自定义模型
- 支持本地部署(vLLM)和API访问
数据集获取
bash python -m PosterAgent.create_dataset
评估方法
-
PaperQuiz评估: bash python -m Paper2Poster-eval.eval_poster_pipeline --metric=qa
-
VLM-as-Judge评估: bash python -m Paper2Poster-eval.eval_poster_pipeline --metric=judge
-
统计指标评估: bash python -m Paper2Poster-eval.eval_poster_pipeline --metric=stats
相关资源
- 论文地址: https://arxiv.org/abs/2505.21497
- 项目主页: https://paper2poster.github.io/
- Hugging Face数据集: https://huggingface.co/datasets/Paper2Poster/Paper2Poster
引用格式
bibtex @misc{pang2025paper2postermultimodalposterautomation, title={Paper2Poster: Towards Multimodal Poster Automation from Scientific Papers}, author={Wei Pang and Kevin Qinghong Lin and Xiangru Jian and Xi He and Philip Torr}, year={2025}, eprint={2505.21497}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2505.21497}, }

- 1Paper2Poster: Towards Multimodal Poster Automation from Scientific Papers滑铁卢大学、新加坡国立大学、牛津大学 · 2025年



