varunagrawal / bbox

Python library for 2D/3D bounding boxes

bbox

bbox a Python library that is intended to ease the use of 2D and 3D bounding boxes in areas such as Object Detection by providing a set of flexible primitives and functions that are intuitive and easy to use out of the box.

Features

2D Bounding Box

Easily work with bounding boxes using a simple class that abstracts and maintains various attributes.

from bbox import BBox2D, XYXY

# x, y, w, h
box = BBox2D([0, 0, 32, 32])

# equivalently, in (x1, y1, x2, y2) (aka two point format), we can use
box = BBox2D([0, 0, 31, 31], mode=XYXY)

print(box.x1, box.y1)  # -> 0 0
print(box.x2, box.y2)  # -> 31 31
print(box.height, box.width)  # -> 32 32

# Syntatic sugar for height and width
print(box.h, box.w)  # -> 32 32

Sequence of 2D bounding boxes

Most tasks involve dealing with multiple bounding boxes. This can also be handled conveniently with the BBox2DList class.

bbl = BBox2DList(np.random.randint(10, 4),
                 mode=XYWH)

The above snippet creates a list of 10 bounding boxes neatly abstracted into a convenient object.

Non-maximum Suppression

Need to perform non-maximum suppression? It is as easy as a single function call.

from bbox.utils import nms

# bbl -> BBox2DList
# scores -> list/ndarray of confidence scores
new_boxes = nms(bbl, scores)

Intersection over Union (Jaccard Index)

The Jaccard Index or IoU is a very useful metric for finding similarities between bounding boxes. bbox provides native support for this.

from bbox.metrics import jaccard_index_2d

box1 = BBox2D([0, 0, 32, 32])
box2 = BBox2D([10, 12, 32, 46])

iou = jaccard_index_2d(box1, box2)

We can even use the Jaccard Index to compute a distance metric between boxes as a distance matrix:

from bbox.metrics import multi_jaccard_index_2d

dist = 1 - multi_jaccard_index_2d(bbl, bbl)

3D Bounding Box

bbox also support 3D bounding boxes, providing convenience methods and attributes for working with them.