This project is currently associated with a manuscript under submission to IEEE Transactions on Circuits and Systems for Video Technology (TCSVT).

At the present stage, the full dataset, model checkpoints, and complete codebase are temporarily not public.

Why the resources are not released yet

To protect unpublished research outcomes during the peer-review stage and to avoid unauthorized reuse of unreleased assets, we are postponing the public release of the following materials until the paper is officially accepted and published:

benchmark data
model checkpoints
inference code
evaluation scripts

Planned Release

Once the paper is accepted and formally published, we will release the complete project resources here as soon as possible, including:

Dataset
Model checkpoints
Inference scripts
Training pipeline
Evaluation tools
Usage documentation

Project Summary

SAGE-VQA studies multilingual text-centric reasoning in realistic video-derived environments, with a focus on low-resource and script-diverse languages.
It is designed to address challenging scenarios such as:

motion blur
dynamic occlusion
temporally evolving text
heterogeneous writing directions and layouts

The framework includes three main components:

Script-Aware Visual Alignment
Hybrid Reward Modeling
Multilingual Policy Shaping

Official Project Repository

For updates and announcements, please visit the official GitHub repository:

GitHub: https://github.com/Shajiu/SAGE-VQA

Release Notice

This page currently serves as an official placeholder and release entry for the SAGE-VQA project.
All files will be uploaded immediately after the review process is completed and the paper is formally published.

Citation

Citation information will be added after publication.

Contact

For research-related questions, please refer to the official GitHub repository or contact the authors through the project page.

Thank you for your patience and understanding.

Downloads last month: 11