Figure 3 : Confusion matrix of the system.

From: Automated home-cage behavioural phenotyping of mice

Figure 3

Confusion matrices evaluated on the doubly annotated 'set B' to compare the agreement between (a) the system and human scoring, (b) human to human scoring and (c) the CleverSys system to human scoring. Each entry (x, y) in the confusion matrix is the probability with which an instance of a behaviour x (along rows) is classified as type y (along column), and which is computed as (number of frames annotated as type x and classified as type y)/(number of frames annotated as type x). As a result, values sum to a value of 1 in each row. The higher probabilities along the diagonal and the lower off-diagonal values indicate successful classification for all categories. Using the annotations made by the 'Annotator group 1' as ground truth, the confusion matrix was obtained for measuring agreement between ground truth (row) with system (computer system), with the 'Annotator group 2' (human) and with baseline software (CleverSys commercial system). For a less cluttered visualization, entries with values less than 0.01 are not shown. The colour bar indicates the percentage agreement, with more intense shades of red indicating agreements close to 100% and lighter shades of blue indicating small percentages of agreement.