Figure 3.
The inclusion of Pattern Recognizer for analyzing the Pattern images enhances the accuracy of the Categorizer
(A and B) Bars showing the test metrics (precisions, recalls, and F1 scores) for three Categorizers: one without Pattern Recognizer, one without Animation Analyzer, and one with both. Test results of datasets from Drosophila larvae (A) and mice (B) are shown. In all three Categorizers, the input shapes and the complexity levels are the same for Pattern Recognizer (8 × 8 × 3, level 1 for larva dataset; 32 × 32 × 3, level 3 for mouse dataset), and Animation Analyzer (8 × 8 × 1 × 15, level 1 for larva; 32 × 32 × 1 × 14, level 3 for mouse). Seven videos for larvae (10 larvae per video) and two videos for mice (one mouse per video) that were not used for generating the training datasets were used to generate the testing data (animations and their paired pattern images). The data with visibly missing body parts (e.g., missing limbs/heads) were excluded (56 pairs, which is 1.76% of the total data generated), because they were caused by occasional, false tracking (often due to changes in the background) and might negatively affect the performance of the Categorizer. Such performance decline is not because of any issues of the Categorizer and might bias the testing of Categorizers. A total of 795 pairs (for larvae) and 2,325 pairs (mice) of data were randomly selected and then were sorted by experimenters into different folders under user-defined behavioral names (building ground truth testing datasets). The ground truth testing datasets were then used to test the categorization metrics of the three Categorizers.