intro

def

segmentation task:

* classifiaction - what is the object
* where is the object

types

Semantic: Pixel (Each) is labeled by its texture and other image related properties.
    
Instance: Pixel is labeled as part of a predefined set of objects. Each object is uniquely identified (Can be counted).
    
Panoptic: Pixel (Each) is labeled by its texture and object.

output

output of L is num of classes;

argmax on each L channel will yield with

score

• by definition is mostly imbalanced

• so the Imbalanced classification Scores can be:

  • Balanced Accuracy.

  • Recall, Precision.

  • Dice / F1.

  • Confusion Matrix

• Object Scores

  • IoU.

  • mAP.

resources

• Understanding Evaluation Metrics in Medical Image Segmentation - https://scribe.rip/d289a373a3f

• Evaluating Image Segmentation Models - https://www.jeremyjordan.me/evaluating-image-segmentation-models/

• Image Segmentation — Choosing the Correct Metric - https://scribe.rip/aa21fd5751af

• miseval: A Metric Library for Medical Image Segmentation EVALuation - frankkramer-lab/miseval

• Kaggle: All the Segmentation Metrics, Understanding Dice Coefficient, Visual Guide To Understanding Segmentation Metrics - https://www.kaggle.com/code/yassinealouini/all-the-segmentation-metrics

The Loss Function

• Cross Entropy Loss.

• Cross Entropy Loss + Label Smoothing.

• Balanced Cross Entropy / Focal Loss.

• Gradient Friendly Region / Boundary Loss.

resources

• Loss Functions for Image Segmentation - JunMa11/SegLossOdyssey

• 3 Common Loss Functions for Image Segmentation https://dev.to/_aadidev/3-common-loss-functions-for-image-segmentation-545o

• Instance segmentation loss functions - https://softwaremill.com/instance-segmentation-loss-functions/

• Focal Loss: An Efficient Way of Handling Class Imbalance - https://scribe.rip/4855ae1db4cb