Read the summary and the most important questions on CS4065

• Lecture 1 (Oud deel)

  This is a preview. There are 5 more flashcards available for chapter 25/04/2017
  Show more cards here

• What parts does a traditional relevance assessment model consist of? Give two famous retrieval models.

  Vector-space model, learning to rank

• What two techniques are often used to index documents for queries:

  1. TF - Term frequency (#occurences)
  2. IDF - Inverse document frequency (#occurences / #documents that contain this word)

• Give the formulas for precision and for recall. When do you prefer precision over recall? And when recall over precision?

  :
  Precision -> Authentication. Face recognition for example.
  Recall -> Financial transactions, legal records.   
  
  Usually precision is prefered by big search engines.

• What are the categories of textual metadata generally available in a multimedia data collection?

  
  
  
  
  • Dedicated annotations (e.g. tags, professional labels)
  • Texts on websites where multimedia items appear
  • Texts as output of communication via social media (tags, comments) 

• Give 2 techniques to improve the usefulness of available metadata.

  1. Help the query to match metadata (search for new tags appearing in images sharing the original tags).
  2. Classify tags by counting the number of the visual neighbors.

• What are the main bottlenecks in relying on visual analysis of images and video to improve search?

  1. Representative power of features needs imrovement
  2. High sensitivity to variations in images' capture conditions

• Lecture 2 (Oud deel)

  This is a preview. There are 7 more flashcards available for chapter 09/05/2017
  Show more cards here

• What feature representation algorithm is given in the lecture?

  SIFT (Scale-Invariant Feature Transform) - Consists out of the following four steps:
  

  1. Scale-space extrema detection
  2. Keypoint localization
  3. Orientation assignment
  4. Keypoint descriptor 
     

• What technique was presented to help format codeword dictionaries?

  k-means clustering.

• Calculate the following histogram attributes that can be used to compare histograms:

  Intersection calculates simillarity (higher = better), rest calculates dissimilarity.

• What alternative was presented for DFT to represent audio?

  Chroma: