TU Wien:Computer Vision VU (Sablatnig)/Prüfung 2021-03-12

Image Classification and Machine Learning

True/False Questions

• On average, images of man-made environments have more vertical and horizontal edges than images of natural environments.
• The bias and variance of a classifier are independent of each other.
• One way of adding spatial information to the bag of words model is spatial pyramid matching.
• Regularization in machine learning means to reduce the number of training samples.
• Training of neural networks is only possible with linear activation functions.
• The main advantage of the bag of words model is that it does not need training images

Sift

(A picture from the SIFT-Algorithm from the slides is given)
The following picture illustrates the process of computing SIFT keypoint descriptors.

• What is shown on the left? What do the green grid, black arrows and blue circle mean?
• The descriptor is illustrated on the right. What do the green grid and black arrows mean? What is the dimensionality of this descriptor?

3D Reconstruction

• What is Structure from Motion?
• Given a fundamental matrix relating the images 1 and 2 and a point x in image 1, what can be said about the corresponding point in image 2?

Image Acquistion

• Describe the principle of lenses (sketch) and the "thin lens" law!
• Continuing from the question above, explain the Depth of Field
• What are the internal camera parameters (sketch) and what influence do they have?

Plenoptic function

• What is described by the plenoptic function?
• Why does a sheet of paper in a scene not reflect an image of the scene? What simple construction can be used to let an image of a scene appear on a sheet of paper?

Image Pyramids

• How are Gaussian Pyramids generated and what is their frequency composition on the different pyramid levels?
• Following from the question above, what is the frequency composition of a Laplacian pyramid?

RANSAC/Image Stitching

• Describe how RANSAC can be used to determine the correct matches and the transformation between two views for image stitching

Scene Understandin

• What are Scene Emergent Features?
• What information is contained in the magnitudes of a Fourier spectrum and how can this information be used to differentiate between man-made and natural scenes?

Harris Corner Detection

• How is the "cornerness" of an image pixel computed by the Harris corner detection algorithm?
• How are the final corner points detected from the cornerness?

Deep Learning

• What is the difference between the "classical" machine learning pipeline for image recognition and the new deep learning methods?
• Explain the concept and usage of auto-encoders.