Accuracy of automated tracking in comparison to human observations.
M. persicae feeding behaviour was measured on Arabidopsis leaf discs by automated video tracking and human observations simultaneously. (a) Out of 139 probes of 16 aphids scored by hand, 88% was detected by automated video tracking. Probes were considered a match when their duration overlapped at least partially. Some of the detected probes were matched by too few (underrated) or too many (overrated) probes. For these situations, the amount of missed or redundant probes is shown. 17 Probes went undetected and 20 false probes were recorded. Mean duration per probe is shown above the bars. (b) Screenshots of the top-view video used for automated tracking. The lower image (σ) shows an aphid positioned on the edge of the leaf disc for more than 20 min, causing overrated probe counts by the software due to continuous switching between an “on the leaf disc” and “off the leaf disc” status. (c) Differences between software and human observations per matched probe. 95% Confidence Intervals are shown above the histograms. Negative values correspond to too early probe starts, too early probe endings, resp. too short duration of probes compared to the human observations. In case of overrated probe counts, the probe with the most similar duration as the manually scored probe was included. The outlier caused by the example in (b) is annotated with σ.