Understanding Dice Loss for Crisp Boundary Detection

Abstraction: Dice loss as alternative to cross-entropy for object boundary detection tasks

Key points:

• Boundary detection is an ill-posed problem; predictions often include spurious boundaries from textures and irrelevant structures
• Dice loss is presented as a far better alternative to cross-entropy loss specifically for boundary detection tasks
• Task focuses only on object geometry — colors, textures, and illuminations are irrelevant
• Standard cross-entropy treats foreground/background pixels with class imbalance typical in boundary maps
• Dice loss directly optimizes the overlap coefficient between predicted and ground-truth boundary regions

Connections: Loss Functions · Image Segmentation · Computer Vision

Source: https://medium.com/ai-salon/understanding-dice-loss-for-crisp-boundary-detection-bb30c2e5f62b