VALORANT Round Outcome Prediction
Collaborated with NAIST to build a transformer-based approach for predicting VALORANT round outcomes from minimap video analysis.
Open External LinkOverview
This project focuses on predicting the outcome of rounds in the tactical shooter VALORANT using deep learning. By analyzing minimap video data, the model forecasts which team will win a round with high accuracy.
The system treats the minimap as a dense summary of player movement and team coordination. That made it possible to build a robust computer-vision pipeline without needing privileged in-game telemetry.
Key Features
Transformer Architecture
Used TimeSformer to capture temporal and spatial dependencies in minimap video sequences.
Large-Scale Dataset
Built a dataset from 1,376 tournament videos to train and validate the model.
Reliable Predictions
Reached 80.55% accuracy in forecasting round outcomes from live match state.
Video-First Pipeline
Processed minimap footage to infer player positions and team momentum in real time.
Technologies Used
Outcomes & Impact
- Published at IEEE Conference on Games 2025.
- Demonstrated a novel use of video transformers in esports analytics.
- Created a reusable dataset for follow-on research.
Collaboration
Conducted in collaboration with the Nara Institute of Science and Technology (NAIST).