This project supports buildling datasets for downstream tasks using many open-source text-spotting datasets. It includes a number of classes and tools to quickly and easily build text spotting datasets with many annotation types, including
- Dot
- 2-point bounding boxes
- Quadrilateral bounding boxes
- Polygons
- Bezier Curves
Additionally, this library makes converting between types extremely easy through an intuitive and extensible interface.
Use pip to install:
pip install textmark
TextAnnotation
is a base class that can be easily extended to support other annotation formats. This library includes several formats already. Of particular note is that TextAnnotation includes a class level conversion registry. Subclasses can be registered like so:
# Add your subclass name:
TextAnnotation.register_name(name, "My_Class")
# Add any relevant conversions
# you only need to convert to the "closest" next annotation type
TextAnnotation.register_conversion(BoxAnnotation, QuadAnnotation, BoxAnnotation.to_quad)
TextAnnotation.register_conversion(QuadAnnotation, BoxAnnotation, QuadAnnotation.to_box)
Now, a user can easily convert between Box and Quad annotations through the use of
my_quad_annotation.to("Box")
If there are additional registries, such as Quad <--> Polygon, the user can convert all the way from a Box to a Polygon annotation (or vice versa) in a single command:
my_polygon_annotation.to("Box")
This system works by constructing a graph of all registered conversions. When a user calls the .to
method, the graph is searched for the target class, and then applies all conversions on that path. This library implements the following simple conversions, which can be automatically chained together:
Polygon <--> Quad
Quad <--> Box
Box <--> Dot
Note that moving up through the chain is a lossy process!!
In addition, Bezier Curves can be converted to Polygons.
This library also includes an easy visualization system. See the example below:
from textmark import TextAnnotation, Visualizer
img_path = ...
my_annotation = TextAnnotation.factory(
"Poly", scene_text, language, *list_of_points
)
my_second_annotation = TextAnnotation.factory(
...
)
# Can visualize an arbitrary number of annotations
vis = Visualizer(img_path, [my_annotation, my_second_annotation, ...])
visualization = vis.visualize()
visualize.show() # uses PIL