Image Processing Reference
In-Depth Information
Table 7
Evaluation of the Chase Classifier Performance on Nontraining Videos
Algorithms
Accuracy Sensitivity Specificity
JAABA
0.826
0.949
0.764
GradientBoost 0.801
0.989
0.750
Logistic
0.871
0.987
0.766
lSVM
0.880
0.915
0.839
gSVM
0.762
0.932
0.668
The accuracy, sensitivity, and specificity scores are shown for each.
Intriguingly, we found gross differences in fly behavior between the videos, which generally
relected the sex ratio composition of their respective experiments. It is precisely differences
such as these that we hope we will be able to detect using methods such those we have deve-
loped here, so this is an encouraging result. A higher proportion of females results in a greater
likelihood of observing females on patch, as well as an increased prevalence of Multifly blobs
(
Table 8
). There was a trend towards more frequent aggression with an increase in male num-
bers, while courtship was observed more frequently at intermediate sex ratios. These results
are entirely consistent with what we understand about fly behavior, which again provides en-
couragement regarding the usefulness of automated annotation of fly behavior such as that
we present here.
Table 8
Behavioral Profiles of the Five Fly Videos (v1-v5), with the Number of Females:Males in the
Experiment (but not Necessarily in Frame) Shown in Brackets
Behaviors
v1(15:5) v2(15:5) v3(10:10) v4(10:10) v5(5:15)
Multifly %
32.61
28.09
15.43
22.46
17.23
Sex = 0%
51.34
63.54
22.75
44.72
24.28
Aggression % 0.29
0.987
1.40
1.73
1.45
Courtship %
0.66
0.915
0.82
1.65
0.73
The percentage of frames containing at least one of these behaviors is shown. Behavioral annotations were ob-
tained using JAABA.
4 Discussion
In this paper we have developed a method for generating, and error correcting, tracking
data from video recordings of
Drosophila
made in nonideal conditions. Nonoptimal conditions
cause problems at the initial image processing stage, due to poor performance of background
subtraction routines, occlusion caused by proximity between animals, and uncertainty in the
number of objects that need to be tracked (
c
.
f
. [
1
]). This leads to subsequent poor performance
Search WWH ::
Custom Search