In this paper, we delve into the crucial role of visual and textual representations in gaming. Traditional Visual and Text Encoders often fall short in this domain. To address this, we introduce a novel multimodal framework, gCLIP, to enhance game understanding. Our approach involves a unique dataset of game screens with human-labeled captions and an adaptation of the Contrastive Language–Image Pretraining (CLIP) model using soft-labels. This method is tailored for nuanced game content analysis. Applications of our methodology include zero-shot classification, game genre clustering, game screen classification, and player risk assessment, showcasing the versatility and impact of our approach. Our study aims to bridge a gap in game-related AI research, offering insights for game developers, players, and researchers.
This repository contains the code necessary for training the gCLIP model, as well as documentation and examples.
- Install Required Modules:
pip install -r requirements.txt
- Download the Dataset:
- Obtain the gCLIP Dataset.
- Place the downloaded data in the
./datadirectory.
- Train the gCLIP Model:
python -m src.train
Hanwool Lee
Jonghyun Choi
Sungbum Jung
- The model and dataset for "gCLIP: A Multimodal Approach to Integrating Visual and Textual Cues in Games," is scheduled for release in January 2024.
- This framework is aimed at enhancing the understanding and interaction within game environments.
- [Paper] Learning Transferable Visual Models From Natural Language Supervision
- [Repo] moein-shariatnia/OpenAI-CLIP
- [Repo] openai/CLIP
- [Repo] Distributed Training in PyTorch
If you find our work useful in your research, please consider citing:
@misc{gCLIP,
author = {Hanwool Lee and Jonghyun Choi and Sungbum Jung},
title = {gCLIP: A Multimodal Approach to Integrating Visual and Textual Cues in Games},
year = {2024},
howpublished = {\url{https://github.com/finMU/gCLIP}}}