search

MhLiao / TextBoxes_plusplus

TextBoxes++: A Single-Shot Oriented Scene Text Detector
Other
956 stars 281 forks source link

Training is not convergent！ #108

Open JentMan opened 5 years ago

JentMan commented 5 years ago

hi，I used the nature scene pictures to train the textboxes++ model, but in the training, the mbox_loss changes between 1 and 4, I feel it is so big, so what i can do to reduce the mbox_loss?

MhLiao commented 5 years ago

Some suggestions (1) Check your ground truth. Ensure that the ground truth of the images is correct. (2) Adjust your batch size and learning rate. Empirically, a larger batch size or smaller learning rate may be helpful. (3) Turn off the random crop data augmentation if the above two suggestions do not help.

JentMan commented 5 years ago

@MhLiao, Did you adopt the random crop data augmentation method in your training code? I don't do that for my data set alone. If I want to detect chinese text line, can you recommend some data set to me?