2D Bounding Box Formats

This section describes the different formats of 2D bounding box annotations. These annotation formats are different ways of representing the same annotation and position of the bounding boxes surrounding the objects in an image.

Annotation Formats

Currently, there are three recognized formats:

1. YOLO
2. PASCAL VOC
3. COCO

YOLO Format

The YOLO format describes the bounding boxes in the following way: class xc yc width height.

The class is an integer which represents the object class ID. The object class ID represents the index of the object in a unique set of labels.

For example, if the unique set of labels is the following below.

background
person
car

If the object class ID is 1, then the object label would be "person". Similarly, 2 would point to "car" and 0 is "background".

Consider the following image below as a reference with a bounding box around the person in the center of the image.

YOLO Format

The coordinate xc represents the center of the bounding box normalized to the image width, W. This means that if x_center is the x-coordinate of the center of the bounding box in pixels, then

xc = x_center/W
xc = 646/1280 = 0.5046875

The coordinate yc has the same idea, except that this coordinate is normalized to the image height, H. This means that if y_center is the y-coordinate of the center of the bounding box in pixels, then

yc = y_center/H
yc = 403/720 = 0.5597222

The width is not the width of the image. The width is the normalized width of the bounding box. This means that if bbx_width represents the width of the bounding box in pixels, then

width = bbx_width/W
width = 188/1280 = 0.146875

The height has the same idea, except that this dimension is normalized to the height of the image. This means that if bbx_height represents the height of the bounding box in pixels, then

height = bbx_height/H
height = 460/720 = 0.6388889

Note

The values for xc, yc, width, and height are floating-point values.

Finally, a text file annotation in a Darknet dataset would contain the line 1 0.5046875 0.5597222 0.146875 0.6388889.

PascalVOC Format

The PASCAL VOC format describes the bounding boxes in the following way: class x1 y1 x2 y2.

The class follows the same idea as the YOLO format above. However, the coordinates are represented differently.

Consider the following image below as a reference with a bounding box around the person in the center of the image.

PascalVOC Format

The coordinates point to the corners of the bounding box in pixels as shown below.

Box Corners

If the width of the image is W, and the height of the image is H, then the coordinates in PascalVOC format are described below.

x1 = Xmin/W = 552/1280 = 0.43125
y1 = Ymin/H = 174/720 = 0.241667
x2 = Xmax/W = 740/1280 = 0.578125
y2 = Ymax/H = 634/720 = 0.880556

Note

The coordinates x1, y1, x2, and y2 are all floating-point values.

Finally, a text file annotation in a Darknet dataset would contain the line 1 0.43125 0.241667 0.578125 0.880556.

COCO Format

The COCO format is a combination of both YOLO and PascalVOC format and describes the bounding boxes in the following way: class x1 y1 width height.

The class follows the same idea as the YOLO format. In addition, x1 and y1 follow the same calculations as the x1 and y1 in PascalVOC format. Finally, the width and the height follow the same calculations as the YOLO format as these represent the normalized width and the height of the bounding box respectively.

Consider the following image below as a reference with a bounding box around the person in the center of the image.

COCO Format

x1 = Xmin/W = 552/1280 = 0.43125
y1 = Ymin/H = 174/720 = 0.241667
width = bbx_width/W = 188/1280 = 0.146875
height = bbx_height/H = 460/720 = 0.638889

Note

The coordinates x1, y1, x2, and y2 are all floating-point values.

Finally, a text file annotation in a Darknet dataset would contain the line 1 0.43125 0.241667 0.146875 0.638889

Further Reading

This section has described three different annotation formats for describing 2D bounding box annotations: YOLO, PascalVOC, and COCO.

Next, take a look at the conventions followed for the EdgeFirst Dataset Structure which describes how the file structure is organized depending if the dataset is sequence-based or not.