Semantic Segmentation Architectures
Until state-of-the-art networks for segmentation:
1. FCN
2. LinkNet
3. UNet
4. SegNet
5. DeepLab
Evaluation Metrics
Dice coefficient: intersection over Union-like metric.
where,
p_iis prediction for pixeli;y_iis ground truth for pixeli;Nis the total number of pixels on the image.
Loss for Semantic Segmentation
Soft-Dice Loss:
The ground truth annotation can be represented as a sum of true positives and false negatives of every class.
The predicted pixels are a sum of true and false positives across all classes.
How to turn DC metric into a loss function?
Dice coefficient compares two discrete masks and, therefore, it is a discrete function. To make a loss function out of it, we need to come up with a differentiable function. So, instead of thresholded values like 0 and 1, we can make floating point probabilities in the range of [0, 1]. The function that can help us doing so is a negative logarithm. (Just recall classification cross-entropy loss which also uses negative logarithm for the same reasons).
