• murphy-gaussians: http://ais.informatik.uni-freiburg.de/teaching/ws17/mapping/pdf/murphy-gaussians.pdf

• [Masterpraktikum, Deep Learning Lab][33], Uni Freiburg, WS ²⁰¹⁸⁄₁₉

  • code at github: https://github.com/aisrobots/dl-lab-2018

• Deep Learning Course, Control Section, Uni Freiburg, WS16/17

  • code at github: https://github.com/mllfreiburg/dl_lab_2016

• CS231n: Convolutional Neural Networks for Visual Recognition, Uni Stanford, 2018

• a book: Neural Networks and Deep Learning

  • free online html

• CS224n: Natural Language Processing with Deep Learning, stanford, 2018

• The Projective Camera, note that the pinhole model is only a special class of the projective camera model

• The Perspective Camera, i.e., the pinhole model

• The Weak-Perspective Camera, it is similar to the pinhole camera model except that it groups objects in the space with similar depth and replace their depth with the same value $z$. Thus, during the projection, $\frac{f}{z}$ is the same for all pixels, which is similar to orthographic projection. When to use this model: when the object is far away from the camera such that $z >> f$ and the field of view is small. Parallel lines are preserved during projection.

• The orthographic camera

Camera Calibration

Refer to http://www.cse.iitd.ernet.in/~suban/vision/geometry/node39.html

• Tsai camera model and calibration
• Camera calibration and absolute conic
• Camera calibration and absolute conic
• What does calibration give?, angle between rays, normal vector of a plane through the camera center.
• The image of the absolute conic
• A simple calibration device, using absolute conic for calibration. It needs 5 image of circular points, which are obtained from 3 planes.
• using vanishing points to determine the absolute conic, refer to the lecture slide here

In the orthographic camera model, every object has the same magnification; while in the weak perspective model, distant object looks smaller. Objects have similar $z$ have the same magnification factor.

Weak perspective projection can be considered as a combination of perspective and orthographic projection.

Refer to Simplified Camera Projection Models, pdf.

Projective Geometry

• ideal points, points at infinity
• line at infinity, plane at infinity
• circular points, absolute conic
• vanishing points
  • if a point lies on the plane at infinity, then its image is called vanishing point
  • two parallel lines intersect at a point on the plane at infinity, so the image of the intersection is a vanishing point
  • it can be used for camera calibration! (1) identify two vanishing points; (2) identify the angle of the two corresponding lines; (3) we get a constraint; (4) the intrinsics have 5 degree of freedom, so we need to find five pairs of vanishing points
  • by vanishing points, we can use just one image to calibrate an image!
• vanishing line
  • it is the image of the plane at infinity
  • all vanishing points lie on the vanishing line
• image of the absolute conic (IAC)
  • it is useful for camera calibration!

todo

adaboost, graphical models, SVM face recognition (pattern matching)

boosting gbdt xgb lgb, auc area, roc curve, light gbm random forest, xgboost

crf (conditional random field)

word2vec, lstm

CS230: Deep Learning

Single View Metrology and this assignment and refer to this one