Objective Kata Scores From Two Camera Views.
Research on scoring karate kata objectively, using two synchronised camera views, pose estimation and a deep learning model trained on expert scores.
Key numbers
0.802
rank correlation with expert judges' scores
0.141
mean squared error on performances it had never seen
~5×
less error than the same model with one camera
50
recorded performances, expanded to 2,400 training samples
What was getting in the way
Karate kata are graded by human judges, which can be subjective and hard to scale. No system existed that produced a single, objective quality score for a whole kata performance.
What we built
A scoring system that watches a kata from the front and the side at the same time, and gives it a quality score that closely tracks expert judges. It was tested on Taikyoku Shodan, the first kata taught in the Shōrin-ryū style.
How it works
- Two cameras. Each performance is filmed from the front and the side, so depth and angles are clearer.
- Pose estimation. MediaPipe tracks the body, and the clearest view is chosen frame by frame.
- Joint angles over time. Elbows, shoulders, hips and knees become a timeline of movement.
- A trained scorer. A deep learning model, trained on scores from two expert karate masters, predicts the final score.
What’s next
First sorting performances into “good” and “needs work”, recognising which kata is being performed, and extending the system to more kata.
Research facts
- Type
- Research Project
- Authors
- Bevin Benet et al.
- Field
- Computer vision · Sport
- Tools
- MediaPipe, Python, Bidirectional LSTM
Where this expertise applies
Custom AI Systems & Fine-Tuned Models
Build AI-powered applications around your specific processes, information and requirements.
Your workflow could be next
Have a Problem Worth Solving?
Tell us what your team keeps doing manually. We’ll help identify what can be improved and how to validate the results before a wider rollout.