Understanding the wide and height label in a bounding box

[Camilo_Vargas]

Hi,
I was watching one of the videos and how b_h and b_w are defined during labeling.

In the image above , as said in the video, b_h is 0.3 and b_w is 0.4. This means that the width of the object is 30% of the width of the bounding box and 40% of the height of the bounding box. But what about in the image below (cat)?

If I want to label row 3 column 2. What would be the values of b_h and b_w? Is b_h and b_w equal to 1.3? I don’t understand how a box is labelled when the target image is bigger than the box itself.

Thank you.
Best course ever!

[ai_curious]

@Camilo_Vargas

This means that the width of the object is 30% of the width of the bounding box

Not quite. First, b_w and b_h are the shape of the bounding box, not the object. In this case it is a ground truth bounding box, I think (not a predicted one). In the first image, the one of the truck, 0.3 and 0.4 are ratios of the ground truth bounding box shape to the entire image shape.

In the second image, which has added the concept of grid cells a la YOLO, the values of b_w and b_h represent ratios of the ground truth bounding box to the grid cell shape. b_w = 1.0 is thus a bounding box exactly the same width as one grid cell. 1.3 means 30% larger than one grid cell.

There is a much more elaborate examination of YOLO grid cell and bounding box relationships in these previous threads:

Applying YOLO anchor boxes

Applying YOLO Anchor Boxes A previous thread provides an introduction to extracting from labelled ground truth data a small set, on the order of 3 <= k <= 10 , of shapes for what is referred to in the YOLO 9000 paper as dimension clusters or priors, but in this class as anchor boxes . NOTE: Just to make things more confusing, the YOLO 9000 paper uses anchor boxes to refer instead to the hand-picked shapes described in the Faster R-CNN paper, not the ones generated by K-Means. The YOLO 9000 pap…

Deep dive on the YOLO v2 localization predictions

YOLO v2 makes 3 types of predictions on an image region. Is there an object there at all? What is it? Where is it? This thread examines the latter, the localization predictions. There are five sets of numbers that play a role in these predictions; everything from converting 3^{rd} party training data to visualizing predicted bounding boxes on an image. There is some flexibility in the which set of numbers is used where, but they have to all be aligned and used consistently. Here are the p…