Intersection over Union (IoU)

It is a Quality Measure used in Image Segmentation.

Visually, it can be represented as follows.

Using the Pixels that fall in the intersection between FP, TP and FN pixel classifications, an accurate Image Segmentation can be obtained.

Intersection over union ($\text{IoU}$)

$$\boxed{\mathrm{IoU}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}+\mathrm{FP}}}$$

• Simply measures overlap between predicted segmentation and ground truth
• Problem: Large objects have more pixels, so they dominate the score

Instance-level intersection over union ($\text{iIoU}$)

$$\boxed{\mathrm{iIoU}=\frac{\mathrm{iTP}}{\mathrm{iTP}+\mathrm{iFN}+\mathrm{FP}}}$$

• Modifies IoU to treat objects more fairly regardless of their size
• Weights each object by: (average object size) / (this object’s size)
• Makes segmenting small objects just as important as large ones