Open czh0001 opened 3 years ago
@czh0001 No, we did not use the gt from task2.
Thank you for your answering! I mean I need to train the character segmentation mask which let the loss_char_mask not to be zero. And I found that the 'use_charann' in 'maskrcnn_benchmark.config.paths_catalog' is the key to train the character segmentation mask when our character-level annotation is already prepared. When the value of 'use_charann' is ‘True’, it seems that the training of the character segmentation mask is processed.
I have trained on the task4(end to end) of icdar2015. And now I need to train on the task2(text segmentation) of icdar2013 but I don't know how to input the datasets.