Blank output ! about ocrsegment

@watersink can you reply back, can you train a working model?

[lvaleriu]

@mrocr Hello. I'm trying to train the model on the uw3 dataset too, but I get a different issue: Incompatible shapes: [1,128,128,1] vs. [1,825,648,1].

Would it be possible to exchange on skype/slack etc on this issue?

Thanks!

@lvaleriu Ocrsegment seems to be incomplete, my suggestion for you is to search other projects.

[lvaleriu]

I've just fixed the incompatible shapes issue (had to modify the resize parameters of images/labels) and trained the network for 8K iterations and got a nice result:

@lvaleriu send the original image, so I could compare your result.

[lvaleriu]

The image is ./make_training_labels/W001.png.

@lvaleriu can you share your latest training results, how many iterations have you reached?

Though I would still urge you to use the original project https://github.com/NVlabs/ocropus3 it contains ocroseg which can also extract the lines.