image segmentation

• references:
  • https://www.youtube.com/watch?v=AZr64OxshLo
  • https://www.youtube.com/watch?v=0FmNxqLFeYo
• questions :

Types of image segmentation

Semantic segmentation

Instance segmentation

Apporaches for segmentation

Using Classical Approach

Histogram Based Segmentation

thresholding

Random Walker Segmentation

Watershed Segmentation

Graph based segmentation

• </home/felladog/Downloads/CV_Nptel/W2/DL4CV_Week02_Part05.pdf>
• https://youtu.be/0HbRnFTOOms?t=680

Using Unsupervised Approach

Gaussian Mixture Model

Kmeans Clustering

Using Deep Learning

UNET

Image Segmentation Loss/Evaluation Method

• Problem with other common metrics
• We need metrics that target pixels in foreground

Intersection over union (Jaccard index)

• sklearn jaccard module

• Overall accuracy is not a good metric for semantic segmentation, especially for imbalanced dataset

• IOU = \(\frac{Intersection}{Union}\)

  = \(\frac{\sum{\hat{y}*y}} {\sum{\hat{y} + y - \hat{y} * y}}\)

  = \(\frac{TP}{TP + FP + FN}\)

Dice Score

• Easy to differentiate
• For Image = \(\frac{2*\sum{\hat{y}*y}} {\sum{\hat{y} + y}}\)
• priortise both precision and recall equally
• Loss = 1 - Dice

IOU vs Dice Score

IoU metric tends to penalize single instances of bad classification more than the F1/Dice score quantitatively even when they can both agree that this one instance is bad.¹