Thanks for the repo. CRAFT works really well with the test data I have. Without using RefineNet, we are able to generate word-level texts, which is what we need before feeding the texts into the Recognition module.
Because with the technique in this repo we cannot really recognize space if we feed a long text to the network. So we are thinking to group a chunk of same-level texts in an array to resolve the space issue.
Most of the texts in our test data are heavily rotated. We are thinking to calculate the average rotational degree of each bounding boxes, recursively feed it to the network and rotate the picture until it is not-skewed. That way, I think we can roughly group together texts that are supposedly in a sentence.
However, I do feel like this is a rather computational heavy. Do you guys have a better approach?
Thanks for the repo. CRAFT works really well with the test data I have. Without using RefineNet, we are able to generate word-level texts, which is what we need before feeding the texts into the Recognition module.
Because with the technique in this repo we cannot really recognize space if we feed a long text to the network. So we are thinking to group a chunk of same-level texts in an array to resolve the space issue.
Most of the texts in our test data are heavily rotated. We are thinking to calculate the average rotational degree of each bounding boxes, recursively feed it to the network and rotate the picture until it is
not-skewed
. That way, I think we can roughly group together texts that are supposedly in a sentence.However, I do feel like this is a rather computational heavy. Do you guys have a better approach?