Figure 5.

Comparison of SparkMaster 2 (SM2) and human annotators using synthetic data. A, An example of synthetic data produced, with a synthetic image to the left and a mask of ground truth spark segmentation to the right. B, The comparison of the performance of the 3 annotation systems. SparkMaster used thresholds of 3.4 and 3.8 for the more- and less-sensitive version respectively. SM2 used the thresholds of 0.45 and 0.6 for the more- and less-sensitive annotation (the threshold values are not comparable between SparkMaster and SM2). The total number of false negatives and false positives across 30 images in the dataset is shown.