Skip to content
choi-yeeunPublic

About

[CVPRW'25] A Training-Free, Task-Agnostic Framework for Enhancing MLLM Performance on High-Resolution Images

Resources

Stars

9 stars

Watchers

0 watching

Forks

Latest commit

 

History

9 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 

Repository files navigation

A Training-Free, Task-Agnostic Framework for Enhancing MLLM Performance on High-Resolution Images

Jaeseong Lee*, Yeeun Choi*, Heechan Choi*, Hanjung Kim, Seon Joo Kim

CVPR 2025 Workshop Emergent Visual Abilities and Limits of Foundation Models (EVAL-FoMo 2) accepted!

ECP

🐳 Environment Setup (Docker)

To simplify deployment and ensure consistent environments, we recommend using the official Qwen2.5-VL Docker image.

📝 Reference: Qwen2.5-VL GitHub Repository

📦 Launch the container

docker run --gpus all --ipc=host --network=host --rm --name qwen2.5 -it qwenllm/qwenvl:2.5-cu121 bash

This launches a container with all necessary dependencies pre-installed, based on Qwen2.5-VL's official setup.


⚡ Quick Start

Step 1: Dataset Preparation

Clone this repo and download datasets as follows:

# Clone repository
git clone https://github.com/yeencye/ECP.git

📂 GUI-grounding task

Inside GUI-grounding/, download the ScreenSpot-Pro dataset

📂 MLLM-perception task

Inside MLLM-perception/, download the HR-Bench dataset


Step 2: 1st-Stage EC (Extract Candidate)

Run the following script to generate candidate regions:

bash GUI-grounding/EC.sh

This will save cropped regions and intermediate JSONs under ./EC_output.


Step 3: 2nd-Stage P (Predict)

Run the following script to refine the bounding boxes using Qwen2.5VL or OS-Atlas:

bash GUI-grounding/P.sh

This will run inference on the cropped regions and save final prediction results under ./P_output.


About

[CVPRW'25] A Training-Free, Task-Agnostic Framework for Enhancing MLLM Performance on High-Resolution Images

Resources

Stars

9 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages