Because bees are small and occupy a small percent of the image, the
distribution of classes considered for detection is in fact skewed in favor of the background.
Detection results on the annotated dataset are summarized in Figure 4. As shown, the simpler background subtraction
based method vastly outperforms the cascade classification
based algorithm on the test sequence.
The end result of applying object tracking is to achieve a
count of how many bees have entered and exited the hive.
Performance for arrival rate and departure rate are measured in
terms of percent error from their respective ground truth. In
early tests, the system over counts departures by 18% and over
counts arrivals by 44% on the annotated test sequence. It is
hard to judge performance based on this metric, however. If
multiple detections occur for a single bee, they are each
tracked. Thus, if a bee which is being tracked more than once
crosses the entrance boundary, it counts for multiple arrivals or
departures.