Skip to content

Light weight toolkit for bounding boxes providing conversion between bounding box types and simple computations.

License

Notifications You must be signed in to change notification settings

devrimcavusoglu/pybboxes

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

72 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

PyBboxes

Python versions Total downloads Monthly downloads
Python versions Build Python versions

Light weight toolkit for bounding boxes providing conversion between bounding box types and simple computations. Supported bounding box types (italicized text indicates normalized values):

  • albumentations : Albumentations Format
    • [x-tl, y-tl, x-br, y-br] (Normalized VOC Format) Top-left coordinates & Bottom-right coordinates
  • coco : COCO (Common Objects in Context)
    • [x-tl, y-tl, w, h] Top-left corner & width & height
  • fiftyone : FiftyOne
    • [x-tl, y-tl, w, h] (Normalized COCO Format) Top-left coordinates & width & height
  • voc : Pascal VOC
    • [x-tl, y-tl, x-br, y-br] Top-left coordinates & Bottom-right coordinates
  • yolo : YOLO
    • [x-c, y-c, w, h] Center coordinates & width & height

Glossary

  • tl: top-left
  • br: bottom-right
  • h: height
  • w: width
  • c: center

News 🔥

  • 2024/10/07 - Annotations are supported for YOLO, COCO and VOC formats.

Roadmap 🛣️

  • Annotation file support.
  • (Upcoming) 3D Bounding Box support.
  • (Upcoming) Polygon support.

Important Notice

Support for Python<3.8 will be dropped starting version 0.2 though the development for Python3.6 and Python3.7 may continue where it will be developed under version 0.1.x for future versions. This may introduce; however, certain discrepancies and/or unsupported operations in the 0.1.x versions. To fully utilize and benefit from the entire package, we recommend using Python3.8 at minimum (Python>=3.8).

Installation

Through pip (recommended),

pip install pybboxes

or build from source,

git clone https://github.com/devrimcavusoglu/pybboxes.git
cd pybboxes
python setup.py install

Bounding Boxes

You can easily create bounding box as easy as

from pybboxes import BoundingBox

my_coco_box = [98, 345, 322, 117]
coco_bbox = BoundingBox.from_coco(*my_coco_box)  # <[98 345 322 117] (322x117) | Image: (?x?)>
# or alternatively
# coco_bbox = BoundingBox.from_array(my_coco_box)

Out of Bounds Boxes

Pybboxes supports OOB boxes, there exists a keyword strict in both Box classes (construction) and in functional modules. When strict=True, it does not allow out-of-bounds boxes to be constructed and raises an exception, while it does allow out-of-bounds boxes to be constructed and used when strict=False. Also, there is a property is_oob that indicates whether a particular bouding box is OOB or not.

Important Note that, if the return value for is_oob is None, then it indicates that OOB status is unknown (e.g. image size required to determine, but not given). Thus, values None and False indicates different information.

from pybboxes import BoundingBox

image_size = (640, 480)
my_coco_box = [98, 345, 580, 245]  # OOB box for 640x480
coco_bbox = BoundingBox.from_coco(*my_coco_box, image_size=image_size)  # Exception
# ValueError: Given bounding box values is out of bounds. To silently skip out of bounds cases pass 'strict=False'.

coco_bbox = BoundingBox.from_coco(*my_coco_box, image_size=image_size, strict=False)  # No Exception
coco_bbox.is_oob  # True

If you want to allow OOB, but still check OOB status, you should use strict=False and is_oob where needed.

Conversion

With the BoundingBox class the conversion is as easy as one method call.

from pybboxes import BoundingBox

my_coco_box = [98, 345, 322, 117]
coco_bbox = BoundingBox.from_coco(*my_coco_box)  # <[98 345 322 117] (322x117) | Image: (?x?)>
voc_bbox = coco_bbox.to_voc()  # <[98 345 420 462] (322x117) | Image: (?x?)>
voc_bbox_values = coco_bbox.to_voc(return_values=True)  # (98, 345, 420, 462)

However, if you try to make conversion between two bounding boxes that require scaling/normalization it'll give an error

from pybboxes import BoundingBox

my_coco_box = [98, 345, 322, 117]
coco_bbox = BoundingBox.from_coco(*my_coco_box)  # <[98 345 322 117] (322x117) | Image: (?x?)>
# yolo_bbox = coco_bbox.to_yolo()  # this will raise an exception

# You need to set image_size for coco_bbox and then you're good to go
coco_bbox.image_size = (640, 480)
yolo_bbox = coco_bbox.to_yolo()  # <[0.4047 0.8406 0.5031 0.2437] (322x117) | Image: (640x480)>

Image size associated with the bounding box can be given at the instantiation or while using classmethods e.g from_coco().

from pybboxes import BoundingBox

my_coco_box = [98, 345, 322, 117]
coco_bbox = BoundingBox.from_coco(*my_coco_box, image_size=(640, 480))  # <[98 345 322 117] (322x117) | Image: (640x480)>
# no longer raises exception
yolo_bbox = coco_bbox.to_yolo()  # <[0.4047 0.8406 0.5031 0.2437] (322x117) | Image: (640x480)> 

Box operations

Box operations now available as of v0.1.0.

from pybboxes import BoundingBox

my_coco_box = [98, 345, 322, 117]
my_coco_box2 = [90, 350, 310, 122]
coco_bbox = BoundingBox.from_coco(*my_coco_box, image_size=(640, 480))
coco_bbox2 = BoundingBox.from_coco(*my_coco_box2, image_size=(640, 480))

iou = coco_bbox.iou(coco_bbox2)  # 0.8117110631149508
area_union = coco_bbox + coco_bbox2  # 41670 | alternative way: coco_bbox.union(coco_bbox2)
total_area = coco_bbox.area + coco_bbox2.area  # 75494  (not union)
intersection_area = coco_bbox * coco_bbox2  # 33824 | alternative way: coco_bbox.intersection(coco_bbox2)
first_bbox_diff = coco_bbox - coco_bbox2  # 3850
second_bbox_diff = coco_bbox2 - coco_bbox  # 3996
bbox_ratio = coco_bbox / coco_bbox2 # 0.9961396086726599 (not IOU)

Functional

Note: functional computations are moved under pybboxes.functional starting with the version 0.1.0. The only exception is that convert_bbox() which still can be used by importing pybboxes only (for backward compatibility).

Conversion

You are able to convert from any bounding box type to another.

import pybboxes as pbx

coco_bbox = (1,2,3,4)  # COCO Format bbox as (x-tl,y-tl,w,h)
voc_bbox = (1,2,3,4)  # Pascal VOC Format bbox as (x-tl,y-tl,x-br,y-br)
pbx.convert_bbox(coco_bbox, from_type="coco", to_type="voc")  # (1, 2, 4, 6)
pbx.convert_bbox(voc_bbox, from_type="voc", to_type="coco")  # (1, 2, 2, 2)

Some formats require image width and height information for scaling, e.g. YOLO bbox (resulting coordinates are rounded to 2 decimals to ease reading).

import pybboxes as pbx

voc_bbox = (1,2,3,4)  # Pascal VOC Format bbox as (x-tl,y-tl,x-br,y-br)
pbx.convert_bbox(voc_bbox, from_type="voc", to_type="yolo", image_size=(28, 28))  # (0.07, 0.11, 0.07, 0.07)

Computation

You can also make computations on supported bounding box formats.

import pybboxes.functional as pbf

coco_bbox = (1,2,3,4)  # COCO Format bbox as (x-tl,y-tl,w,h)
voc_bbox = (1,2,3,4)  # Pascal VOC Format bbox as (x-tl,y-tl,x-br,y-br)
pbf.compute_area(coco_bbox, bbox_type="coco")  # 12
pbf.compute_area(voc_bbox, bbox_type="voc")  # 4

Annotation file conversion

pybboxes now supports the conversion of annotation file(s) across different annotation formats. (yolo, voc and coco are currently supported)

This is a 3 step process.

1. Instantiate the Annotations class

from pybboxes.annotations import Annotations

anns = Annotations(annotation_type='yolo')

Important you have to explicitly declare annotation_type beforehand. post declaration, you will be only able to load annotation in declared format but you will be able to export to other annotation formats.

2. Load the annotations file

After you have instantiated the Annotations class declaring annotation_type, you can now load the annotations using appropriate method of the Annotations class.

2.1 Load from yolo

from pybboxes.annotations import Annotations

anns = Annotations(annotation_type='yolo')

anns.load_from_yolo(labels_dir='./labels', images_dir='./images', classes_file='./classes.txt')

As yolo normalizes the bounding box metadata, path to corresponding images directory must be provided (via images_dir) so that physical dimension of image data can be inferred.

Also, path to classes_file (usually classes.txt) should be provided that lists all the class labels that is used for the annotation. Without this, pybboxes will fail to assign appropriate class labels when converting across different annotations format.

2.2 Load from voc

from pybboxes.annotations import Annotations

anns = Annotations(annotation_type='voc')

anns.load_from_voc(labels_dir='./labels')

2.3 Load from coco

from pybboxes.annotations import Annotations

anns = Annotations(annotation_type='coco')

anns.load_from_coco(json_path='./validation.json')

3. Saving annotations to different format

3.1 Saving annotations to yolo format

As every image data has its own corresponding annotation file in yolo format, you have to provide path to export_dir where all the annotation files will be written.

from pybboxes.annotations import Annotations

anns = Annotations(annotation_type='coco') # just for the demonstration purpose

anns.load_from_coco(json_path='./validation.json') # we could have loaded the annotation data from other format as well

anns.save_as_yolo(export_dir='./labels')

This will create all the required annotation files (in yolo format) in given directory. Additionally, it will also create classes.txt in the given folder which will list all the class labels used for the annotation.

3.2 Saving annotations to voc format

Just like yolo format, in voc format, every image data has also its own corresponding annotation file. So, you have to provide path to export_dir where all the annotation files will be written.

from pybboxes.annotations import Annotations

anns = Annotations(annotation_type='coco') # just for the demonstration purpose

anns.load_from_coco(json_path='./validation.json') # we could have loaded the annotation data from other format as well

anns.save_as_voc(export_dir='./labels')

3.3 Saving annotations to coco format

To export annotations in coco format, you just have to provide name (or path) of the output file (in json format) via export_file.

from pybboxes.annotations import Annotations

anns = Annotations(annotation_type='voc') # just for the demonstration purpose

anns.load_from_voc(labels_dir='./labels') # we could have loaded the annotation data from other format as well

anns.save_as_coco(export_file='./validation.json')

Contributing

Installation

Install the package as follows, which will set you ready for the development mode.

pip install -e .[dev]

Tests

To tests simply run.

python -m tests.run_tests

Code Style

To check code style,

python -m tests.run_code_style check

To format codebase,

python -m tests.run_code_style format

License

Licensed under the MIT License.

About

Light weight toolkit for bounding boxes providing conversion between bounding box types and simple computations.

Topics

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages