EDIR
收藏资源简介:
EDIR是由阿里巴巴通义实验室等机构联合构建的细粒度组合图像检索基准数据集,旨在解决现有基准类别覆盖不足和模态偏差问题。该数据集包含5000条结构化查询,涵盖5个主类别和15个子类别,数据来源通过图像编辑技术合成,确保修改类型的精确控制和多样性。其构建流程包括种子图像筛选、原始三元组生成、查询重写及人工验证,最终应用于评估多模态嵌入模型在复杂真实场景下的组合推理能力,尤其针对电商、创意搜索等领域的细粒度需求。
EDIR is a fine-grained compositional image retrieval benchmark dataset jointly constructed by Alibaba's Tongyi Lab and other institutions, designed to address the problems of insufficient category coverage and modal bias in existing benchmarks. This dataset includes 5,000 structured queries covering 5 main categories and 15 sub-categories. Its data is synthesized via image editing technologies, ensuring precise control and diversity of modification types. The construction pipeline of EDIR comprises seed image screening, original triplet generation, query rewriting and manual verification, and it is ultimately applied to evaluate the compositional reasoning capabilities of multi-modal embedding models in complex real-world scenarios, especially targeting the fine-grained demands of fields such as e-commerce and creative search.
EDIR 数据集概述
数据集基本信息
- 数据集名称:EDIR
- 官方论文:Rethinking Composed Image Retrieval Evaluation: A Fine-Grained Benchmark from Image Editing
- 论文链接:https://arxiv.org/abs/2601.16125
- 论文状态:ACL 2025
- 数据集状态:将于未来两个月内在Hugging Face平台发布
数据集用途
- 核心任务:组合图像检索评估
- 特点:一个从图像编辑任务中衍生出的细粒度基准数据集
数据集获取与使用
-
获取方式:需通过指定链接下载至
dataset/edir目录 -
评估脚本:使用
main.py脚本进行评估 -
评估命令示例: sh python main.py --model_id "rzen-7b" --model_name_or_path "" --dataset edir --dataset_path [图像路径]
-
自定义模型评估:参考
models目录实现自有模型 -
自定义数据集添加:参考
test目录的数据格式
引用信息
如需引用本工作,请使用以下BibTeX条目: bibtex @misc{song2026edir, title={Rethinking Composed Image Retrieval Evaluation: A Fine-Grained Benchmark from Image Editing}, author={Tingyu Song and Yanzhao Zhang and Mingxin Li and Zhuoning Guo and Dingkun Long and Pengjun Xie and Siyue Zhang and Yilun Zhao and Shu Wu}, year={2026}, eprint={2601.16125}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2601.16125}, }




