TU Wien:Computer Vision VU (Sablatnig)/Prüfung 2019-03-08

Aus VoWi
Zur Navigation springen Zur Suche springen

90 Minuten Zeit, nur eine Gruppe, Taschenrechner für letzte Frage erforderlich

Camera True/False Kreuzerl (-1 Point for wrong answers)[Bearbeiten | Quelltext bearbeiten]

  • In contrast to lenses the pinhole camera projects an orthographic image
  • Perspective Projection is a non-linear transform
  • Parallel lines remain parallel in perspective projection
  • Camera measures part of plenoptic function
  • Focal length is a extrinsic camera parameter
  • Texture is a depth cue

Machine Learning[Bearbeiten | Quelltext bearbeiten]

  • Supervised vs. Unsupervised
  • Generalization
  • Gradient Descent

Harris[Bearbeiten | Quelltext bearbeiten]

Describe how constructed

Gaussian[Bearbeiten | Quelltext bearbeiten]

  • Describe what is Gaussian Pyramid and how is it constructed
  • Describe how Laplacian Pyramid can be constructed from

SIFT[Bearbeiten | Quelltext bearbeiten]

  • Describe how Sift is made scale-invariant?
  • Is vl_dsift scale-invariant, argument why?

Epilar Definitions[Bearbeiten | Quelltext bearbeiten]

  • Baseline
  • Epipol
  • Epipolar line
  • Disparity
  • Fundamental Matrix

Stereo True/False Kreuzerl (-1 Point for wrong answers)[Bearbeiten | Quelltext bearbeiten]

  • In Structure-from-Motion the location of the 3D Scene Point and the Camera Point is calculated
  • BRDF is given per material
  • A Trifocal-Tensor is a relationship between 3 views of the scene
  • With P given the 3D Scene Point can be reconstructed
  • Disparity is a binocular depth cue
  • Essential Matrix is 3x3

Bag of Words[Bearbeiten | Quelltext bearbeiten]

  • How is codeword dictionary generated?
  • What is representation of image for classification?

Gabor[Bearbeiten | Quelltext bearbeiten]

  • What are Gabor Wavelets?
  • What is the Real and Imaginary Component?

RANSAC[Bearbeiten | Quelltext bearbeiten]

  • 300 correct matchings in 500 features, what is the probability of drawing 4 inliers?
  • How many runs are necessary to construct a correct homography with 99% confidence.
  • Why is it good to construct the final homography from all inliers of the best run?