Open michath opened 8 years ago
Hi, so far we have tested several variants:
for ICDAR dataset, the ROC curve does not improve much (but it can be profitable for specific task)
do you have suggestion?
only that threshold optimization should probably account for overlapping between correct and incorrect text regions: precision matters more while overlapping whereas recall matters more while there is no overlapping.
seems the local threshold in the paper is implemented as global (=13) in tools/segmentation.py: does it indeed? if so, how much effort is required to make it locally adaptive? thanks!