Semi-automatic Segmentation & Alignment of Handwritten

Intersection-over-Union-based alignment

Intersection-over-Union-based alignment is a reasonable approach to aligning bounding boxes to each other and is the chosen method when aligning using GT coordinates. The method calculates the Intersection over Union (IoU) between each segmented bounding box and each GT bounding box. For each GT box, the segmented bounding box with the greatest value gets connected to the string content attached to the GT box. IoU is calculated as the intersection area between two boxes divided by the union area of the two boxes (Figure 10). Thus, the method factors in how similar boxes are, in terms of size and offset to each other. Although, IoU-based alignment also suffers from the same volatility that the distance-based approach does. Figure 11 showcases what different IoU values might look like when comparing to the GT.

Figure 10: Visual representation of Intersection over Union between two boxes, used in the calculation for IoU

1. 
   (b)
   

(c)

Figure 11: An example of what an IoU value of (a) 0.934, (b) 0.477 and (c) 0.219 looks like. (a) and (b) are from IAM, green boxes are the segmentation from the algorithm; the red boxes are the GT. (c) is from Labour’s Memory, the green boxes are the segmentation from the algorithm; all the other colours are the GT.

Figure 12: The structure of the XML file including the layout structure information, the ”CON- TENT” attribute contains the text string and coordinates of the bounding box.

4.