Releases: bertsky/ocrd_detectron2
Releases · bertsky/ocrd_detectron2
v0.1.8
Fixed
- workarounds for broken models (DocBank_X101, Jambo-sudo_X101)
make deps
: add explicit reqs prior to pip step with Torch index
- set
pc:PcGts/@pcGtsId
from mets:file/@ID
Added
- CI for CLI tests (with cached models and stored result artifacts)
Changed
- migrated model URLs from external to Github release assets
v0.1.7
Fixed
- adapt to Numpy 1.24 (no
np.bool
)
Added
- model by Jambo-sudo (PubLayNet+custom GT)
- model by LayoutParser (PRImA Layout GT)
- CLI tests
v0.1.6
Fixed
- avoid colon in generated region IDs
make deps
: add explicit deps for torch
- fix/update file resources
- fix model config base paths on-the-fly
Added
v0.1.5
Fixed
- param
debug_img
: 1 image per page
- URLs/specs for PubLayNet/JPLeoRX models
v0.1.4
Added
- param
postprocessing
(select steps, including none
)
- param
debug_img
(styles to visualise raw predictions, including none
)
v0.1.3
Fixed
make deps
: fall back to Detectron2 src build
Changed
- added various models as file resources
- added corresponding preset files
- updated documentation
v0.1.2
Fixed
make deps
: fix CUDA detection even more
- apply
device
param as passed
Changed
- downscale images to no more than 150 DPI for prediction (for speed)
- add param
operation_level
(default page
), add table
mode
v0.1.1
Fixed
make deps
: fix CUDA detection and allow CPU as fallback
Changed
- instance segmentation postprocessing: use asymmetric overlap
criterion for non-maximum suppression
- skip instances which belong to classes with empty category
- annotate incrementally (by skipping candidates that overlap
with pre-existing top-level regions)