Image Processing Reference
In-Depth Information
Back Versus Dykes Versus Smart—All Metrics
Technique
Correctness (%)
Unweighted Points 71
Weighted Points 68.1
Euclidean Distance 57.1
6.7 Selecting the Best Metrics
If all 45 pieces were to be used to train the system, the resulting classifier data would represent
what data from a randomized trial would look like on average. In this case, one can see that
Bach's chord inversion statistics are far different from those of Dykes and Smart. Bach also re-
lies more heavily on secondary chords:
Classifier Data from 45 Test Pieces
Five Metrics
Bach Dykes Smart
Root Position (%)
65.7 62
61.1
First Inversion (%)
22.1 20
22.02
Second Inversion (%)
2
10.72
10
Third Inversion (%)
1.1
0.6
1.9
Secondary Chords (%) 11.5 4.4
3.9
To test if FPC could even more accurately distinguish between Bach and either of the others,
20 additional trials were run for Bach and Dykes using only root position, first inversion,
second inversion, third inversion, and secondary chords metrics.
Bach Versus Dykes—Five Metrics
Technique
Correctness (%)
Unweighted Points 80.7
Weighted Points 88.6
Euclidean Distance 89.3
Although Unweighted Points was 1.8% points less accurate, Weighted Points improved by
1.8% points, and Euclidean Distance was a surprising 17.8% points more accurate. Whereas
Euclidean Distance performed the worst last time, this time it actually performed the best.
Using only these five metrics likely removed considerable amounts of “noisy” data, which
suggests Euclidean Distance performs best with low noise.
 
Search WWH ::




Custom Search