qurator-spk/eynollah
1
0
Fork 0
mirror of https://github.com/qurator-spk/eynollah.git synced 2026-03-02 05:11:57 +01:00
Code Issues Projects Releases Wiki Activity

docs/train: document --missing-printspace=project

Browse source
This commit is contained in:
Robert Sachunsky 2026-02-25 13:18:40 +01:00
parent 4202a1b2db
commit 42bab0f935
2 changed files with 10 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

5
docs/train.md
View file

@ -271,6 +271,11 @@ eynollah-training generate-gt pagexml2label \
-doi "dir of output cropped images" -doi "dir of output cropped images"
``` ```
Also, note that it can be detrimental to layout training if there are visible segments which
the annotation does not account for (and thus the model must learn to ignore). So if the images
are not cropped, the `-ps` _should_ be used. If a PAGE XML file is missing `PrintSpace` (or `Border`)
annotations, use `-mps` to either `skip` these or `project` (i.e. crop from existing segments).
## Train a model ## Train a model
### classification ### classification

6
src/eynollah/training/gt_gen_utils.py
View file

@ -724,7 +724,11 @@ def get_images_of_ground_truth(
print(gt_list[index], "has no Border or PrintSpace - projecting hull of segments") print(gt_list[index], "has no Border or PrintSpace - projecting hull of segments")
bboxes = list(map(bbox_from_points, coords)) bboxes = list(map(bbox_from_points, coords))
left, top, right, bottom = zip(*bboxes) left, top, right, bottom = zip(*bboxes)
ps_bbox = [min(left), min(top), max(right), max(bottom)] left = max(0, min(left) - 5)
top = max(0, min(top) - 5)
right = min(x_len, max(right) + 5)
bottom = min(y_len, max(bottom) + 5)
ps_bbox = [left, top, right, bottom]
else: else:
print(gt_list[index], "has no Border or PrintSpace - using full page") print(gt_list[index], "has no Border or PrintSpace - using full page")
ps_bbox = [0, 0, None, None] ps_bbox = [0, 0, None, None]