What are IoU / Jaccard and Tversky losses?
hardAnswer
- IoU (Jaccard) = |P ∩ G| / |P ∪ G|; loss = 1 - IoU.
- Similar to Dice but stricter.
- Tversky loss generalizes: T(P, G) = |P∩G| / , where α, β control the penalty on false positives vs false negatives.
- Handy for skewed segmentation where you want to trade recall vs precision. α=β=0.5 recovers Dice.
Check yourself — multiple choice
- IoU and Dice are identical
- IoU = intersection/union (stricter than Dice); Tversky adds tunable FN/FP weights
- Only for classification
- Tversky ignores overlap
IoU: overlap/union; Tversky: tunable FN/FP trade-off, generalizes Dice.
#losses#segmentation
Practise Deep Learning
214 interview questions in this topic.
Related questions
- Write Dice loss and its role in segmentation.
- Why compute softmax + cross-entropy jointly via log-sum-exp?
- How should the output head be designed for a regression task with a strictly positive target?
- How is the output head different for multi-label vs multi-class classification?
- What is label smoothing and why does it help?
- What is a dilated (atrous) convolution?