05 - Advanced Segmentation

Question 1

Motivation for advanced segmentation

Answer 1

• artefacts
• noise
• gradients
• noise & unsharpness

Question 2

Automated thresholding methods (unsupervised)

Answer 2

• histogram-based (Otsu using derivatives, Intermodes using midpoint b/ 2 peaks, Isodata)
• image-based (like entropy)
• result-based (bias?)

Question 3

Hysteresis thresholding

Answer 3

• when one threshold is too hign and the other too small

- strict threshold, remove small obj, loose threshold, combine, keep connected pixels

Question 4

Feature vectors

Answer 4

• pairing spatial info with other info (value)
• K-mean clustering (distance metric => cluster)
• – random centers, group by dist, recalc center, iterate
• quad-trees (split until criterion fulfilled in subregion)
• superpixels (groups of similar pixels)
• probabilistic models of segmentation

Question 5

Supervised segmentation approaches

Answer 5

• workflow: train - predict
• K-nearest neighbors
• linear regression
• trees, forest (combine results from trees)

Question 6

Warnings

Answer 6

• loss function should be carefully chosen
• train != validate
• optimizer improvement
• metrics
• over-fitting (can’t generalize)