Closed ghost closed 4 years ago
@deepseek The picture you gave can't be displayed, whether the training is normal or not, and whether loss drops normally or not.
@deepseek Have you tested the model I provided, and can it detect the text normally?
I chaged the im_scale
to 1000, now getting somewhat better detection, but still ctpn have alot of downfalls
@BADBADBADBOY When training on a text-line dataset, download here: dataset download ctpn_2100.pth It can't detect the text-lines correctly, even after training for 2100 epochs. Waiting for your reply
Sample results: