README.md

---
title: ReasonMatch
emoji: 🧩
colorFrom: blue
colorTo: purple
sdk: static
pinned: false
---

# ReasonMatch

Official Hugging Face organization for **ReasonMatch-Bench** and the paper:

**Eliciting Complex Spatial Reasoning in MLLMs through Wide-Baseline Matching**

[![Paper](https://img.shields.io/badge/arXiv-2606.03577-b31b1b)](https://arxiv.org/abs/2606.03577)
[![Project Page](https://img.shields.io/badge/Project-Page-2458d3)](https://aim-uofa.github.io/reasonmatch/)
[![Code](https://img.shields.io/badge/GitHub-Code-111111)](https://github.com/aim-uofa/ReasonMatch)
[![Dataset](https://img.shields.io/badge/HuggingFace-Dataset-ffcc4d)](https://huggingface.co/datasets/ReasonMatch/ReasonMatch)

## Overview

**ReasonMatch-Bench** evaluates visual reasoning over point correspondences. It focuses on **wide-baseline matching**, where models must reason across viewpoint changes, geometric transformations, fine-grained visual evidence, and occlusion.

The project also includes **Dynamic Correspondence Reinforcement Learning (DCRL)**, a verifiable reinforcement learning recipe for wide-baseline matching without explicit chain-of-thought supervision.

## Links

| Resource | Link |
|---|---|
| Paper | https://arxiv.org/abs/2606.03577 |
| Project page | https://aim-uofa.github.io/reasonmatch/ |
| Code | https://github.com/aim-uofa/ReasonMatch |
| Dataset | https://huggingface.co/datasets/ReasonMatch/ReasonMatch |
| Original ModelScope dataset | https://www.modelscope.cn/datasets/jxzh2020/ReasonMatchBench |
| Pretrained weights | To be announced |

## Resources

- **ReasonMatch-Bench**: in-domain evaluation benchmark for visual matching.
- **Out-of-domain evaluation suite**: additional generalization evaluation set.
- **Training recipe**: veRL-based DCRL training code for multimodal models.
- **Evaluation code**: scripts for ReasonMatch-Bench and out-of-domain evaluation.

## Dataset

The public evaluation archives are hosted on Hugging Face:

```bash
pip install -U huggingface_hub

hf download ReasonMatch/ReasonMatch \
  reasonmatch_bench.tar.gz ood_dataset.tar.gz \
  --repo-type dataset --local-dir ReasonMatchBench

tar -xzf ReasonMatchBench/reasonmatch_bench.tar.gz -C ReasonMatchBench
tar -xzf ReasonMatchBench/ood_dataset.tar.gz -C ReasonMatchBench
```

They are also available from the original ModelScope dataset:

```bash
dataset_path=/your/absolute/path/to/datasets
dataset_repo=jxzh2020/ReasonMatchBench

modelscope download --repo-type dataset "${dataset_repo}" \
  --include 'reasonmatch_bench.tar.gz' 'ood_dataset.tar.gz' \
  --local_dir "${dataset_path}"

tar -xzf "${dataset_path}/reasonmatch_bench.tar.gz" -C "${dataset_path}"
tar -xzf "${dataset_path}/ood_dataset.tar.gz" -C "${dataset_path}"
```

After extraction:

- `reasonmatch_bench/` contains ReasonMatch-Bench and `annotation_index.json`.
- `ood_dataset/` contains the out-of-domain evaluation set.

## Citation

```bibtex
@InProceedings{Zhong_2026_CVPR,
    author    = {Zhong, Hao and Zhu, Muzhi and Zeng, Shenyan and Li, Anzhou and Chen, Cong and Geng, Hua and Shi, Duochao and Ye, Wentao and Lin, Tao and Chen, Hao and Shen, Chunhua},
    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}
}
```