<a class="user-mention notranslate" data-hovercard-type="user" data-hovercard-url="/us

IIRC Downsample images (1/4th or smaller). Cut up images into

Training the weights.caffemodel about hisdb HOT 3 CLOSED

mrocr commented on August 10, 2024

Training the weights.caffemodel

from hisdb.

ctensmeyer commented on August 10, 2024

We used the Caffe framework using the model arch specified in the repo.

On Thu, Mar 21, 2019 at 2:47 AM mrocr ***@***.***> wrote: @ctensmeyer <https://github.com/ctensmeyer> @herobd <https://github.com/herobd> @cwig <https://github.com/cwig> Thank you for your hard work. For Task 3 - Textline segmentation, how did you train the weights.caffemodel <https://github.com/ctensmeyer/HisDB/blob/master/task_3/weights.caffemodel> ? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#1>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AC2mpOTNljJGAUO29hlc9-59wTfAKJCqks5vY0cXgaJpZM4cA8lQ> .

from hisdb.

mrocr commented on August 10, 2024

So you used your own Caffe fork along with the model structure that you set.

1- How to prepare the training data?
2- How to actually train?
3- How does HisDB performance compared to start_follow_read?

Your help would be appreciated.

from hisdb.

ctensmeyer commented on August 10, 2024

IIRC

Downsample images (1/4th or smaller). Cut up images into overlapping patches (e.g. 256.256) for training inputs. For the GT, render the baselines as 7-pixel thick lines to create a binary mask for each image.
To actually train, I set up a Caffe solver file and a training model.prototxt and then used the caffe train cmd.
SFR probably does better. We haven't done a direct comparison because they output different representations of text lines.

from hisdb.