qurator-spk
/
sbb_pixelwise_segmentation
mirror of https://github.com/qurator-spk/sbb_pixelwise_segmentation.git
1
0
Fork 0
Code Issues Releases Wiki Activity

Merge pull request #7 from qurator-spk/update-readme

Browse Source 
Update README.md
pull/15/head
vahidrezanezhad 5 years ago committed by GitHub
parent 8bdb295cac a9c86b2bcb
commit 5b4df66d86
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 41 additions and 29 deletions
Show Stats Download Patch File Download Diff File

68
README.md
Unescape Escape View File

@ -1,36 +1,48 @@
# Train # Pixelwise Segmentation
just run: python train.py with config_params.json > Pixelwise segmentation for document images
## Introduction
This repository contains the source code for training an encoder model for document image segmentation.
# Ground truth format ## Installation
Either clone the repository via `git clone https://github.com/qurator-spk/sbb_pixelwise_segmentation.git` or download and unpack the [ZIP](https://github.com/qurator-spk/sbb_pixelwise_segmentation/archive/master.zip).
Lables for each pixel is identified by a number . So if you have a ### Pretrained encoder
binary case n_classes should be set to 2 and Download our pretrained weights and add them to a ``pretrained_model`` folder:
labels should be 0 and 1 for each class and pixel. ~~https://file.spk-berlin.de:8443/pretrained_encoder/~~
In the case of multiclass just set n_classes to the number of classes ## Usage
you have and the try to produce the labels
by pixels set from 0 , 1 ,2 .., n_classes-1.
The labels format should be png.
If you have an image label for binary case it should look like this: ### Train
To train a model, run: ``python train.py with config_params.json``
### Ground truth format
Lables for each pixel are identified by a number. So if you have a
binary case, ``n_classes`` should be set to ``2`` and labels should
be ``0`` and ``1`` for each class and pixel.
In the case of multiclass, just set ``n_classes`` to the number of classes
you have and the try to produce the labels by pixels set from ``0 , 1 ,2 .., n_classes-1``.
The labels format should be png.
If you have an image label for a binary case it should look like this:
Label: [ [[1 0 0 1], [1 0 0 1] ,[1 0 0 1]], Label: [ [[1 0 0 1], [1 0 0 1] ,[1 0 0 1]],
[[1 0 0 1], [1 0 0 1] ,[1 0 0 1]] , [[1 0 0 1], [1 0 0 1] ,[1 0 0 1]] ,
[[1 0 0 1], [1 0 0 1] ,[1 0 0 1]] ] [[1 0 0 1], [1 0 0 1] ,[1 0 0 1]] ]
This means that you have an image by 3*4*3 and pixel[0,0] belongs
to class 1 and pixel[0,1] to class 0. This means that you have an image by `3*4*3` and `pixel[0,0]` belongs
to class `1` and `pixel[0,1]` belongs to class `0`.
# Training , evaluation and output
train and evaluation folder should have subfolder of images and labels. ### Training , evaluation and output
And output folder should be empty folder which the output model will be written there. The train and evaluation folders should contain subfolders of images and labels.
The output folder should be an empty folder where the output model will be written to.
# Patches
### Patches
if you want to train your model with patches, the height and width of If you want to train your model with patches, the height and width of
patches should be defined and also number of the patches should be defined and also the number of batches (how many patches
batchs (how many patches should be seen by model by each iteration). should be seen by the model in each iteration).
In the case that model should see the image once, like page extraction,
the patches should be set to false. In the case that the model should see the image once, like page extraction,
# Pretrained encoder patches should be set to ``false``.
Download weights from this link and add it to pretrained_model folder.
https://file.spk-berlin.de:8443/pretrained_encoder/

Write Preview
Loading…
Cancel
Save