Evaluation
Let N the total number of word instances for every query (ground
truth), M the total number of detected word instances and C the
correctly detected word instances. The evaluation will be based on the
metrics of recall (RC), precision (PR) and F-measure (FM) which are
defined as follows:
RC = (100*C)/N %
PR = (100*C)/M %
FM = (2*RC*PR)/(RC+PR)
A word instance is considered as detected only if there is a significant overlap (>80%) with an existing word spotting result. The overlap is expressed by the intersection over union metric
IOU = (A∩B)/(A∪B)
where A and B denote the bounding box areas of the word instance in the ground truth and a detected word, respectively.