Transforms on Rotated Bounding Boxes

This example illustrates how to define and use rotated bounding boxes.

Note

Support for rotated bounding boxes was released in TorchVision 0.23 and is currently a BETA feature. We don’t expect the API to change, but there may be some rare edge-cases. If you find any issues, please report them on our bug tracker: https://github.com/pytorch/vision/issues?q=is:open+is:issue

First, a bit of setup code:

from PIL import Image
from pathlib import Path
import matplotlib.pyplot as plt


import torch
from torchvision.tv_tensors import BoundingBoxes
from torchvision.transforms import v2
from helpers import plot

plt.rcParams["figure.figsize"] = [10, 5]
plt.rcParams["savefig.bbox"] = "tight"

# if you change the seed, make sure that the randomly-applied transforms
# properly show that the image can be both transformed and *not* transformed!
torch.manual_seed(0)

# If you're trying to run that on Colab, you can download the assets and the
# helpers from https://github.com/pytorch/vision/tree/main/gallery/
orig_img = Image.open(Path('../assets') / 'leaning_tower.jpg')

Creating a Rotated Bounding Box

Rotated bounding boxes are created by instantiating the BoundingBoxes class. It’s the format parameter of the constructor that determines if a bounding box is rotated or not. In this instance, we use the CXCYWHR BoundingBoxFormat. The first two values are the X and Y coordinates of the center of the bounding box. The next two values are the width and height of the bounding box, and the last value is the rotation of the bounding box, in degrees.

orig_box = BoundingBoxes(
    [
        [860.0, 1100, 570, 1840, -7],
    ],
    format="CXCYWHR",
    canvas_size=(orig_img.size[1], orig_img.size[0]),
)

plot([(orig_img, orig_box)], bbox_width=10)

Transforms illustrations

Using RandomRotation:

rotater = v2.RandomRotation(degrees=(0, 180), expand=True)
rotated_imgs = [rotater((orig_img, orig_box)) for _ in range(4)]
plot([(orig_img, orig_box)] + rotated_imgs, bbox_width=10)

Using Pad:

padded_imgs_and_boxes = [
    v2.Pad(padding=padding)(orig_img, orig_box)
    for padding in (30, 50, 100, 200)
]
plot([(orig_img, orig_box)] + padded_imgs_and_boxes, bbox_width=10)

Using Resize:

resized_imgs = [
    v2.Resize(size=size)(orig_img, orig_box)
    for size in (30, 50, 100, orig_img.size)
]
plot([(orig_img, orig_box)] + resized_imgs, bbox_width=5)

Note that the bounding box looking bigger in the images with less pixels is an artifact, not reality. That is merely the rasterised representation of the bounding box’s boundaries appearing bigger because we specify a fixed width of that rasterized line. When the image is, say, only 30 pixels wide, a line that is 3 pixels wide is relatively large.

Clamping Mode, and its effect on transforms

Some transforms, such as CenterCrop, may result in having the transformed bounding box partially outside of the transformed (cropped) image. In general, this may happen on most of the geometric transforms.

In such cases, the bounding box is clamped to the transformed image size based on its clamping_mode attribute. There are three values for clamping_mode, which determines how the box is clamped after a transformation:

• None: No clamping is applied, and the bounding box may be partially outside of the image.

• “hard”: The box is clamped to the image size, such that all its corners are within the image canvas. This potentially results in a loss of information, and it can lead to unintuitive resuts. But may be necessary for some applications e.g. if the model doesn’t support boxes outside of their image.

• “soft”: . This is an intermediate mode between None and “hard”: the box is clamped, but not as strictly as in “hard” mode. Some box dimensions may still be outside of the image. This is the default when constucting BoundingBoxes.

Note

For axis-aligned bounding boxes, the “soft” and “hard” modes behave the same, as the bounding box is always clamped to the image size.

Let’s illustrate the clamping modes with CenterCrop transform:

assert orig_box.clamping_mode == "soft"

box_hard_clamping = BoundingBoxes(orig_box, format=orig_box.format, canvas_size=orig_box.canvas_size, clamping_mode="hard")

box_no_clamping = BoundingBoxes(orig_box, format=orig_box.format, canvas_size=orig_box.canvas_size, clamping_mode=None)

crop_sizes = (800, 1200, 2000, orig_img.size)
soft_center_crops_and_boxes = [
    v2.CenterCrop(size=size)(orig_img, orig_box)
    for size in crop_sizes
]

hard_center_crops_and_boxes = [
    v2.CenterCrop(size=size)(orig_img, box_hard_clamping)
    for size in crop_sizes
]

no_clamping_center_crops_and_boxes = [
    v2.CenterCrop(size=size)(orig_img, box_no_clamping)
    for size in crop_sizes
]

plot([[(orig_img, box_hard_clamping)] + hard_center_crops_and_boxes,
      [(orig_img, orig_box)] + soft_center_crops_and_boxes,
      [(orig_img, box_no_clamping)] + no_clamping_center_crops_and_boxes],
     bbox_width=10)

The plot above shows the “hard” clamping mode, “soft” and None, in this order. While “soft” and None result in similar plots, they do not lead to the exact same clamped boxes. The non-clamped boxes will show dimensions that are further away from the image:

print("boxes with soft clamping:")
print(soft_center_crops_and_boxes)
print()
print("boxes with no clamping:")
print(no_clamping_center_crops_and_boxes)

boxes with soft clamping:
[(<PIL.Image.Image image mode=RGB size=800x800 at 0x7F9F7013FCD0>, BoundingBoxes([[478.8188, 400.9185, 570.0000, 874.1443,  -7.0000]], format=BoundingBoxFormat.CXCYWHR, canvas_size=(800, 800), clamping_mode=soft)), (<PIL.Image.Image image mode=RGB size=1200x1200 at 0x7F9F7013FD90>, BoundingBoxes([[ 678.9319,  600.0001,  569.9992, 1278.9989,   -7.0000]], format=BoundingBoxFormat.CXCYWHR, canvas_size=(1200, 1200), clamping_mode=soft)), (<PIL.Image.Image image mode=RGB size=2000x2000 at 0x7F9F7013FE80>, BoundingBoxes([[1089.0000,  918.0000,  570.0001, 1840.0000,   -7.0000]], format=BoundingBoxFormat.CXCYWHR, canvas_size=(2000, 2000), clamping_mode=soft)), (<PIL.Image.Image image mode=RGB size=2364x1542 at 0x7F9F7013FF40>, BoundingBoxes([[1260.9314,  771.0001,  570.0002, 1623.5675,   -7.0000]], format=BoundingBoxFormat.CXCYWHR, canvas_size=(1542, 2364), clamping_mode=soft))]

boxes with no clamping:
[(<PIL.Image.Image image mode=RGB size=800x800 at 0x7F9F6FD225C0>, BoundingBoxes([[ 489.,  318.,  570., 1840.,   -7.]], format=BoundingBoxFormat.CXCYWHR, canvas_size=(800, 800), clamping_mode=None)), (<PIL.Image.Image image mode=RGB size=1200x1200 at 0x7F9F6D166FE0>, BoundingBoxes([[ 689.,  518.,  570., 1840.,   -7.]], format=BoundingBoxFormat.CXCYWHR, canvas_size=(1200, 1200), clamping_mode=None)), (<PIL.Image.Image image mode=RGB size=2000x2000 at 0x7F9F6D166EF0>, BoundingBoxes([[1089.,  918.,  570., 1840.,   -7.]], format=BoundingBoxFormat.CXCYWHR, canvas_size=(2000, 2000), clamping_mode=None)), (<PIL.Image.Image image mode=RGB size=2364x1542 at 0x7F9F701A5B10>, BoundingBoxes([[1271.,  689.,  570., 1840.,   -7.]], format=BoundingBoxFormat.CXCYWHR, canvas_size=(1542, 2364), clamping_mode=None))]

Setting the clamping mode

The clamping_mode attribute, which determines the clamping strategy that is applied to a box, can be set in different ways:

• When constructing the bounding box with its BoundingBoxes constructor, as done in the example above.

• By directly setting the attribute on an existing instance, e.g. boxes.clamping_mode = "hard".

• By calling the SetClampingMode transform.

Also, remember that you can always clamp the bounding box manually by calling the ClampBoundingBoxes() transform! Here’s an example illustrating all of these option:

t = v2.Compose([
    v2.CenterCrop(size=(800,)),  # clamps according to the current clamping_mode
                                 # attribute, in this case set by the constructor
    v2.SetClampingMode(None),  # sets the clamping_mode attribute for future transforms
    v2.Pad(padding=3),  # clamps according to the current clamping_mode
                        # i.e. ``None``
    v2.ClampBoundingBoxes(clamping_mode="soft"),  # clamps with "soft" mode.
])

out_img, out_box = t(orig_img, orig_box)
plot([(orig_img, orig_box), (out_img, out_box)], bbox_width=10)