Segmentation losses & metrics

Before we begin

Segmentation models optimize per-pixel losses and report overlap metrics. Choosing the wrong loss for imbalanced data is a common failure mode.

Learning objectives

Compare cross-entropy, Dice, and focal loss.
Compute IoU and mIoU.
Interpret boundary quality qualitatively.
Handle class imbalance and ignore regions.

Cross-entropy

Standard per-pixel CE — works when classes are roughly balanced. For rare foreground (tumor, defect), CE alone may ignore small regions.

Weighted CE: multiply rare class loss by higher weight.

Dice loss

\text{Dice} = \frac{2|A \cap B|}{|A| + |B|}

Dice loss $= 1 - \text{Dice}$ — directly optimizes overlap; popular in medical imaging.

Focal loss

Down-weights easy pixels — focuses on hard negatives and hard positives. Useful for dense detection/segmentation with extreme imbalance.

mIoU

For each class $c$ :

\text{IoU}_c = \frac{TP_c}{TP_c + FP_c + FN_c}

mIoU = mean of IoU over classes (often excluding void). Primary benchmark metric on Cityscapes, ADE20K, etc.

Qualitative review

Always overlay predictions on hard images — thin structures, shadows, and rare classes often fail before mIoU drops noticeably.

What's next

Module 5 quiz — then pet segmentation project.