Christian Bartz

Hmm, okay... could you provide the exact calls you used to start the `train` and `evaluation` scripts? Maybe I can use that to help you with your problem =)

@wyh410 you are using `--timesteps 1` while evaluating, but you trained with `--timesteps 4` this could be the problem, I guess. @Jacoobr hmm, looks like you changed something in the...

That's very interesting. You seem to have code that is different to the code on github, because the line numbers did not match, as I was checking the code. If...

Hmm, `--timesteps` should be `4` if you trained on the `easy` dataset (because these images always contain four different views, so we need to localize four text regions). Furthermore: Did...

Hmm, there is one more thing I can think of. While looking at the code, I see that the images for training the network are not automatically resized to be...

You can not get any scores for the predicted boxes, at least not with the current implementation (you'd need to add some extra classifiers and think of a way to...

The learning rate is too high! You can check this by having a look at all images in the folder `/bboxes`. You should see that the predicted bboxes jump around...

Do you mean what kind of accuracy is obtained? If so it is word level accuracy. The network does not need many epochs to learn normally. 10 should be enough...

Hmm, two things: 1. it is not necessarily good the set the learning rate to a higher value than `1e-4`, you are using `8e-4` right now. 2. In order to...

It should not come as a surprise to you that the model trained on the FSNS dataset does not work well on other data. This is a common problem in...