ReasonMatch
收藏资源简介：
ReasonMatch-Bench是一个用于评估多模态大语言模型在复杂空间推理能力上的基准数据集。它由论文《通过宽基线匹配激发MLLMs中的复杂空间推理》引入，核心任务是评估模型在宽基线视角变化、几何变换、细粒度视觉证据以及遮挡等挑战下，对图像点对应关系进行视觉推理的能力。数据集包含两个主要部分：主评测集（ReasonMatch-Bench），涵盖dl3dv、re10k、scannet和uco3d等多个源数据集，并包含one_to_one、one_to_multi和multi_to_multi等多种对应关系注释类型；以及域外评测集（OOD evaluation），包含eth3d、imc2022、imc2024和unrealstereo4k等数据集，用于评估模型泛化能力。数据以图像对（view_a, view_b）的形式呈现，每个样本附带有结构化的答案（answer_json）和元数据（metadata_json）。数据集的字段信息完整，包括域、源数据集、注释类型、子集、样本ID、阶段ID、匹配数量、图像数据、答案JSON、元数据JSON、归档文件路径等14个特征。该数据集适用于图像-文本到文本的多模态任务，专注于视觉推理、宽基线匹配和空间推理等研究领域。
ReasonMatch-Bench is a benchmark dataset for evaluating the complex spatial reasoning capabilities of multimodal large language models (MLLMs). It was introduced in the paper titled "Unlocking Complex Spatial Reasoning in MLLMs via Wide-Baseline Matching". Its core task is to evaluate the visual reasoning ability of models on image point correspondences under challenges including wide-baseline viewpoint changes, geometric transformations, fine-grained visual evidence, and occlusion. The dataset consists of two main parts: the main evaluation set (ReasonMatch-Bench) and the out-of-distribution (OOD) evaluation set. The main evaluation set covers multiple source datasets such as dl3dv, re10k, scannet and uco3d, and contains multiple correspondence annotation types such as one_to_one, one_to_multi and multi_to_multi. The OOD evaluation set includes datasets such as eth3d, imc2022, imc2024 and unrealstereo4k, which is designed to evaluate the generalization ability of models. The data is presented in the form of image pairs (view_a, view_b), with each sample accompanied by structured answer_json and metadata_json. The dataset has complete field information, covering 14 features including domain, source dataset, annotation type, subset, sample ID, stage ID, number of matches, image data, answer JSON, metadata JSON, archive file path and other related items. This dataset is applicable to image-text-to-text multimodal tasks, and focuses on research areas including visual reasoning, wide-baseline matching and spatial reasoning.
数据集概述：ReasonMatch
主页：https://aim-uofa.github.io/reasonmatch/ 许可证：Apache-2.0
任务类别：图像-文本到文本 (image-text-to-text)
标签：reasonmatch, visual-reasoning, wide-baseline-matching, multimodal, spatial-reasoning
简介：ReasonMatch-Bench 是一个用于评估多模态大语言模型（MLLMs）在宽基线匹配场景下进行复杂空间推理能力的数据集。其核心任务涉及在存在大幅视角变化、几何变换、细粒度视觉证据和遮挡的情况下，对点对应关系进行视觉推理。该数据集由论文《Eliciting Complex Spatial Reasoning in MLLMs through Wide-Baseline Matching》提出。
数据集结构：
- ReasonMatch-Bench 评估集：
reasonmatch_bench.tar.gz，包含域内评估数据。 - OOD 评估集：
ood_dataset.tar.gz，包含域外评估数据。 - 预览样本：
data/viewer_samples.parquet，包含 40 个代表性样本，仅用于浏览。
预览样本覆盖范围：
- ReasonMatch-Bench 部分：
dl3dv,re10k,scannet,uco3d子集，涵盖one_to_one,one_to_multi,multi_to_multi任务类型。 - OOD 评估部分：
eth3d,imc2022,imc2024,unrealstereo4k子集，涵盖alpha和numeric任务类型。
数据特征：每个样本包含以下字段：
domain：领域source_dataset：源数据集annotation_type：标注类型subset：子集sample_id：样本IDstage_id：阶段IDnum_matches：匹配点数量view_a：视角A图片view_b：视角B图片answer_json：答案（JSON格式）metadata_json：元数据（JSON格式）archive：归档metadata_path：元数据路径view_a_path：视角A图片路径view_b_path：视角B图片路径
下载与使用：
完整评估需要使用 reasonmatch_bench.tar.gz 和 ood_dataset.tar.gz 两个压缩包。可通过 huggingface_hub 工具下载，并解压后获得 reasonmatch_bench/ 和 ood_dataset/ 两个目录。
相关资源：
- 项目页：https://aim-uofa.github.io/reasonmatch/
- 代码库：https://github.com/aim-uofa/ReasonMatch
- 论文：https://arxiv.org/abs/2606.03577
- 原始 ModelScope 数据集：https://www.modelscope.cn/datasets/jxzh2020/ReasonMatchBench
引用：
@InProceedings{Zhong_2026_CVPR, author = {Zhong, Hao and others}, title = {Eliciting Complex Spatial Reasoning in MLLMs through Wide-Baseline Matching}, booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)}, month = {June}, year = {2026}, pages = {16768-16778} }
