README.md

metadata

task_categories:
  - image-text-to-text
license: cc-by-nc-4.0
tags:
  - multimodal
  - llm
  - vision-language
  - visual-reasoning
  - tree-search

ZoomEye: Enhancing Multimodal LLMs with Human-Like Zooming Capabilities through Tree-Based Image Exploration

This repository contains the evaluation data for ZoomEye, a method presented in the paper ZoomEye: Enhancing Multimodal LLMs with Human-Like Zooming Capabilities through Tree-Based Image Exploration.

ZoomEye proposes a training-free, model-agnostic tree search algorithm tailored for vision-level reasoning. It addresses the limitations of existing Multimodal Large Language Models (MLLMs) that operate on fixed visual inputs, especially when dealing with images containing numerous fine-grained elements. By treating an image as a hierarchical tree structure, ZoomEye enables MLLMs to simulate human-like zooming behavior, navigating from root to leaf nodes to gather detailed visual cues necessary for accurate decision-making.

This dataset supports the evaluation of MLLMs on a series of high-resolution benchmarks, demonstrating consistent performance improvements for various models.

• Paper: https://huggingface.co/papers/2411.16044
• Project Page: https://szhanz.github.io/zoomeye/
• Code: https://github.com/om-ai-lab/ZoomEye

Evaluation Data Preparation

The core evaluation data (including V* Bench and HR-Bench) used in the ZoomEye paper has been packaged together.

1. Download Data: The evaluation data is provided here. After downloading, please unzip it. The path to the unzipped data is referred to as <anno path>.

2. [Optional] MME-RealWorld Benchmark: If you wish to evaluate ZoomEye on the MME-RealWorld Benchmark, follow these steps:

   • Follow the instructions in this repository to download the images.
   • Extract the images to the <anno path>/mme-realworld directory.
   • Place the annotation_mme-realworld.json file from this link into <anno path>/mme-realworld.

Folder Tree

The expected folder structure after preparation is as follows:

zoom_eye_data 
  ├── hr-bench_4k                                  
  │   └── annotation_hr-bench_4k.json
  │   └── images/
  │     └── some.jpg
  │    ...
  ├── hr-bench_8k
  │   └── annotation_hr-bench_8k.json
  │   └── images/
  │     └── some.jpg
  │    ...
  ├── vstar
  │   └── annotation_vstar.json
  │   └── direct_attributes/
  │     └── some.jpg
  │    ...
  │   └── relative_positions/
  │     └── some.jpg
  │    ...
  ├── mme-realworld
  │   └── annotation_mme-realworld.json
  │   └── AutonomousDriving/
  │   └── MME-HD-CN/
  │   └── monitoring_images/
  │   └── ocr_cc/
  │   └── remote_sensing/
 ...

Sample Usage

1. Run the Python Demo

We provide a demo file of Zoom Eye accepting any input Image-Question pair. The zoomed views of Zoom Eye will be saved into the demo folder.

python ZoomEye/demo.py \
    --model-path lmms-lab/llava-onevision-qwen2-7b-ov \
    --input_image demo/demo.jpg \
    --question "What is the color of the soda can?"

2. Run the Gradio Demo

We also provide a Gradio Demo. Run the script and open http://127.0.0.1:7860/ in your browser.

python gdemo_gradio.py

Citation

If you find this repository helpful to your research, please cite our paper:

@article{shen2024zoomeye,
  title={ZoomEye: Enhancing Multimodal LLMs with Human-Like Zooming Capabilities through Tree-Based Image Exploration},
  author={Shen, Haozhan and Zhao, Kangjia and Zhao, Tiancheng and Xu, Ruochen and Zhang, Zilun and Zhu, Mingwei and Yin, Jianwei},
  journal={arXiv preprint arXiv:2411.16044},
  year={2024}
}