1 free read left. Create a free account for unlimited reading + AI research.

WhitepaperJournal of Games, Game Art, and Gamification30 Jun 2024

YOLO-based Mobile Legends Match Result Parsing

Key insights

6 takeaways

01
The YOLOv8-based parsing system achieves a 25x increase in processing speed compared to traditional multi-stage CNN methods, reducing inference time from 0.807 seconds to 0.032 seconds per image.See it on page 2
02
The model maintains perfect accuracy and F1-scores while automating the extraction of hero icons, team medals, and match statistics from Mobile Legends screenshots.See it on page 1
03
The framework utilizes a hybrid approach, combining YOLOv8 for object detection in a single forward pass with PaddleOCR for extracting text-based data like battle IDs and scores.See it on page 2
04
Training was conducted over 150 epochs on an Nvidia P100 GPU using a validated dataset of 5,329 screenshots, split into 4,229 training images and 1,100 test images.See it on page 2
05
Operational efficiency is currently limited by the need for manual adjustments to confidence thresholds to mitigate overlapping detections during the parsing process.See it on page 3
06
Future scalability of the system depends on developing automated fine-tuning or template matching techniques to adapt to frequent game updates, such as UI changes and new hero releases.See it on page 3

Summary

This study addresses the scarcity of accessible game data for the popular MOBA title Mobile Legends by developing an automated system to parse match result screenshots. Because the game developer does not provide public data APIs, researchers must rely on crowdsourced screenshots. The primary objective is to replace slower, multi-stage convolutional neural network (CNN) classification methods with a more efficient, unified object detection framework based on YOLOv8.

The methodology involves training a YOLOv8 model on a dataset of 5,329 manually validated match result screenshots. The system performs object detection and classification in a single forward pass, identifying hero icons, team medals, and match statistics. For text-based data, such as battle IDs and scores, the system crops relevant regions for processing via PaddleOCR. The model was trained over 150 epochs using an Nvidia P100 GPU, with data split into 4,229 training images and 1,100 test images. Performance was evaluated using accuracy, F1-score, root mean squared error, and character error rate.

The findings demonstrate that the YOLO-based approach achieves perfect accuracy and F1-scores, matching the performance of the previous CNN-based method while operating 25 times faster. The average inference time per image was reduced from 0.807 seconds to 0.032 seconds. While the model successfully handles complex visual data, the study notes that the system requires manual adjustments to confidence thresholds to prevent overlapping detections. The authors conclude that while this framework significantly improves data collection efficiency, future research should focus on adapting the model to accommodate frequent game updates, such as new hero releases or user interface changes, potentially through template matching or fine-tuning techniques.

YOLO-based Mobile Legends Match Result Parsing

YOLO-based Mobile Legends Match Result Parsing

Key insights

Summary

More from Journal of Games, Game Art, and Gamification

Implementation of a Pathfinding-Based Intelligent Agent for Enemy Behavior in the "Run to BIPOL" Maze Tournament Game

Effective and Immersive Teleoperation with Real-World Constraints

The Gothic Gamification of Diablo: A Framework for the English Composition/First-Year Writing Classroom

Implementation of Finite State Machine in Flowchart-Based Visual Programming Game

Board Game Cafe Analysed: Analysis on Sales and Games at Dhadhu Cafe

Predicting Preference with Sparse Data in Personalized Gamification via Deep Learning

Development of Multilingual Educational Game UI/UX Design for Animal Introduction: Indonesia

Educational Tycoon Game and Cultural Heritage: Developing PAANTU BANG

Developing Mahakarya Vokasi Virtual Reality Game: Player-Centered Design with UEQ-Based Analysis

Gamification, Employee and Emotional Engagement: Leveraging Gamification to Boost Employee Emotional Engagement

Designing and Developing Digital Computer Game of Plastic Waste Awareness for Young Children

User Experience Analysis on Mobile-Based MOBA Games: Mobile Legends

More on AI

TWIG #311 Musk Enters the Game, Archero 2 Aims for the Bullseye, PlayStation Goes Portable

Monster Hunter Now Launch / Luda Raises For AI Sim / CS:GO 2 Launch

Mata: An Android Eye-Tracking Based User Interface Control Application

The Vainglory Story - 5 Lessons From Building a Touch Screen eSport

LVP's Paul Heydon on investment, eSports, mobile MOBAs, and Europe's advantage

Biz Daily: Why Vainglory dev Super Evil raised $26 million

Biz Daily: Can Vainglory overcome the mobile MOBA curse?

TWIG #407: Roblox’s Steal Economy, Tim Sweeney Goes All In on AI and Xbox Loses Its COO

Those ‘decompiling’ videos, iPhone Duo games, Bing Gordon talks EA, Zynga and AI plus plenty more

Why Supercell couldn’t save Squad Busters and Google's AI game maker | Week in Mobile Games podcast

Why Use Tenjin’s MCP Server?

How Riot Went From One Game to Five Hits | Mark Sottosanti, Riot Games