Unsupervised learning and Visual Features

Question 1

How does unsupervised learning work?

Answer 1

Cluster unsorted data based on similarities and patterns

Question 2

How does K-means clustering work?

Answer 2

You want to group the data into K clusters. This is done by minimizing the sum of squares distance between each point and the nearst center.
You start by just guessing centers, and then check for all points. Update with new centers until convergence.

Question 3

How does the Gaussian Mixture model work?

Answer 3

Is used for density estimation. Places a kernel on top of the datapoints (like convolution) and tries to fit the data to this kernel.

Question 4

How does PCA (Principal Component Analysis) work?

Answer 4

PCA is a data compression algorithm (dimensionality reduction). It transfers data to new coordinates, with the greates variance being the first coordinate. Will “lose” some data, but the most important stuff is kept.

Question 5

What are two kinds of Visual Features?

Answer 5

• Hand crafted: You define a math formula and check
  which features fit. Can be split into global and local.
• Learned: Checks the distribution of data, the algorithm
  “learns” what features are in the image.

Question 6

What kinds of descriptors do we have?

Answer 6

• SIFT
• SURF
• HOG
• LBP
• Motion Flow field