Machine learning with Weka for Kung fu- training purposes Victoria Værnø [email protected] vaernoe.wordpress.com
Feb 23, 2016
BackgroundMotion capture and machine learning workbenches are accessible to the public.
Links between cheap hardware (eg. web-cams and Kinect) & open source machine learning software is still hard to find for motion capture.
Research focus
• Build a model which is good enough to consider for a beta application for amateur Kung-Fu training purposes.– For separately created unseen data, high accuracy
and high “bad”-class label precision• What challenges follow a relatively limited
training set and what basic machine learning techniques can reduce these?
Meta-Learner
Clustering
Result- and Data
Analysis
Test
Method
Data
Data attributes: 18 joints * 3 dimensions * 6 frames per movie + 1 class lable = 325 attributes
.bvh
.arff
Iteration 1
Accuracy97%
Precisionof ”Bad”
1
Good! Too good?
Generate MLP model on the training data.
Iteration 1
Accuracy85%
Precisionof ”Bad”
1
Not so great. Unbalanced dataset?
Test the model with unseen dataset.
Class 1 Attribute values which in reality define class 1
Attribute values which in reality define class 2
Attribute values which very often arise in class 1, but does not define class 1.
Unbalanced Data
Class 2
This part is much bigger in the test set and
real life.
Iteration 2
Clustering K-means K=3
Trying to compensate for unbalanced data
• Meta-classifier AdaBoostM1 boosting algorithm• Collecting more data – new people.• Other suggestions, please tell or email me!
Iteration 3
AdaBoostM1
+
Just MLP+ Data set 1
AdaBoostM1+ Data set 1
Just MLP + Data set 1+ Data set 3
AdaBoostM1+ Data set 1+ Data set 3
84%
80%
89%
86%
Generate models and test on Data set 2.
• Promising results for further investigation.– Cross-validation over 90%– ”Bad” class label always precision = 1– Adding just one more person -> unseen data 89%
• Classifying unseen people’s kicks remain unexplored.
• Data collection: Hard for one person to create balanced motion data.
• Modeling: Boosting strategy to combat unbalanced motion capture data works in some cases, but adding a different person’s motion is far more efficient.
”Spend time gathering more data rather than tuning a particular method” Nilsson N.J
Conclusion
Lessons Learned
Thank You!
Q?