1. What are the fiducial (ni isto kot minutiae points) points?
    • Anatomically meaningful landmarks of the face, typically useful for observing growth patterns, such as areas around eyes, lips, eyebrows...
    • Image Upload 1
  2. What is interpupilary distance (IPD)?
    • Razdalja v pikslih med dvema pupiloma, tipično uporabljeno pri normalizaciji.
    • Image Upload 2
  3. What is the IPD for extracting L1 facial features?
    Manj kot 30 IDP
  4. What is the IPD for extracting L2 facial features?
    30 - 75 IPD
  5. What do we observe in L1 facial features?
    • Geometrija obraza
    • Barva kože
    • Spol glede na geometrijo obraza
  6. What do we observe on L2 facial features?
    • Natanča oblika obraza
    • Komponente obraza (Usta, oči, nos), relacija med komponentami
    • Tekstura kože
    • Image Upload 3
  7. What do we observe on L3 facial features?
    • Micro level features on the face
    • Scars
    • Mozolji
    • Mozolji
    • Skin dislocations

    Image Upload 4
  8. What is the rough idea of Adaboost?
    Construct linear combination of weak classifiers where in each iteration we assign higher weights to examples falsely classified in the previous step.
  9. Which are the 2 most effective Haar-like features for face detection?
    Able to reject about 60% of non faces.

    Predel oči je temnejši kot predel lic.

    • Predel nosu svetlejši kot predel oči.
    • Image Upload 5
  10. What is the main problem of geometric warping for facial alignment?
    Geometric warping je postopek kjer poravnamo sliko tako da so oči horizontalno. 

    Problem je da lahko rotitamo le sliko, če ima na sliki glavo obrneno v stran tega pa ne moremo popraviti.

    Image Upload 6
  11. Roughly describe landmark detection using cascaded regression
    Vhod algoritma je slika in pozicije zaznanih obrazov.

    Izhod: Landmarks (fiducial points)
  12. What are 3 main approaches of facial recognition?
    • Appearance based
    • Model based
    • Texture based
  13. What is the idea of appearance-based facial recognition?
    The pixel at the location (x,y) can be expressed as weighted sum of pixel values in all the training images at (x,y).

    • Uporablja intenzitete surovih slikovnih pik.
    • Naredi kompaktno reprenzentacijo celotnega obraza, tako da zmapira high dimensional face image v lower dimensional sub-space. 
    • Ta subspace je definiran kot množica vektorjev, ki so naučeni training set of images.
    • Potem pa uporabimo mapping PCA, LDA in ICA kjer imamo linearne projekcije.
  14. What are 2 main subspace appearance-based approaches?
    • PCA (Principal Component Analysis)
    • LDA (Linear Discriminant analysis)
  15. What is the idea of PCA?
    Uporablja training data to learn a subspace that accounts for as much variability in the training data as possible.

    By performing an Eigen value decomposition of covariance matrix of the data.

    • V glavnem:
    • Izračunam kovarianco matrike
    • Iz te matrike izračunam lastne vrednosti.
  16. What is the idea of LDA?
    Find a suitable projection that maximizes the inner-class difference while minimising the intra-class difference.
  17. What are some problems of PCA?
    Assumes that data has Gausian distribution. Not appropriate for some data sets.
  18. What is the difference between eigenfaces and fisherfaces?
    • Eigenfaces are obtained by PCA
    • Image Upload 7

    • Fisherfaces are obtained by LDA
    • Image Upload 8
  19. What is the idea of model-based facial recognition?
    Try to derive a pose-independent representation based on several fiducial/landmark points.
  20. What is the idea of face bunch graph (FBG) algorithm for model-based fase recognition.
    • Predstavlja obraz kot graf, kjer je vsak node fiducial point.
    • Image Upload 9
  21. What is jet?
Card Set