Trajectory evaluation and behavioral scoring using JAABA in a noisy system - Emerging Trends in Image Processing, Computer Vision, and Pattern Recognition - page 219

ranking of the ML algorithms remained broadly the same in this new data. The gSVM did very
well, and logistic regression did relatively poorly. Again JAABA (GentleBoost) did very well
overall.

Table 3
Evaluation of the ML Algorithm Performance for Multifly Calling on Nontraining Videos

Algorithms Accuracy Sensitivity Specificity

TABU 0.823-0.994 0.747-0.994 0.884-0.995
JAABA 0.986-0.995 0.982-0.994 0.989-0.996
GradientBoost 0.981-0.990 0.979-0.993 0.977-0.987
Logistic 0.983-0.993 0.979-0.992 0.977-0.990

lSVM 0.983-0.992 0.978-0.998 0.976-0.987

gSVM 0.984-0.993 0.975-0.992 0.984-0.993

The minimum to maximum range of accuracy, sensitivity, and specificity scores are shown for each, in comparison
with the input trajectory data (TABU).

Sex Annotation : We found that the various ML algorithms had accuracies in Sex calling ran-
ging between 0.73 and 0.89 ( Table 4 ). Logistic regression, GradientBoost and JAABA all had
similar accuracies, above 0.88. The video blob-processing annotation based on the colored
marker, however, was the most accurate of all the Sex classifiers (with an accuracy of 0.93).
Integrating this classifier together with the JAABA scores improved the performance statistics
for Sex calling to 0.95 and above.

Table 4
Performance Measures of the Sex Classifier on Threefold Cross Validation

Algorithms Accuracy Sensitivity Specificity Precision AUC

blobColor 0.939 0.904 0.975 0.997 -

JAABA 0.883 0.913 0.852 0.860 -

GradientBoost 0.882 0.893 0.871 0.874 0.953

Logistic 0.893 0.867 0.919 0.915 0.952

lSVM 0.816 0.906 0.727 0.769 0.949

gSVM 0.731 0.827 0.635 0.695 0.884

blobColor + JAABA 0.967 0.947 0.987 0.986 -

The accuracy, sensitivity, specificity, precision, and area under the curve scores are shown for each.
On the nontraining data, the performance of the Sex classifier for JAABA was very similar
to that on the trained video (with an accuracy above 0.90), suggesting that this classification is
very robust ( Table 5 ). Indeed, all ML methods however (particularly lSVM) performed some-