Augmentation
#1
by DonkeySmall - opened
I would not use Translation and Mosaic augmentation, because you will get this
Great point!
I’m using the model to detect text in a desktop GUI. Initially I planned to cherry-pick part of the dataset specifically for GUI text detection, but a model trained only on your dataset already performs surprisingly well on this task.
My plan now is to fork the model and maintain a GUI-focused detector by adding dedicated GUI text detection training data, while retraining the core text detector to make it more robust and universal, following your recommendations.