/ComputerVisionExercise

exercise of CV course in TUD

Primary LanguagePython

Computer Vision Exercise

exercise of CV course in TUD

week 10_12 : Neural Network

including :

The result as following :

  • using W of shape[2,2], and discribe this problem as a classification problem, i.e. using softmax_cross_entropy as loss function,( PS: how to draw a line : y = W * x + b with W of shape[2,2], I still don't know. with reference, I got a result as following, but can't explain why there are somewhere are not linear.)

  • using W of shape[2,1], using mean square as loss function

week 7-9 : HDR

including :

The result as following :

  • inversed camera response function: (my implementation used Mitsunagaand Nayar Technique)

  • my result: (I don't know which part got a problem, but I guess it happens during the calculating inversed camera response function. because before I visualize the result of Radiance Map, there are also many black and white dot noise.)

week 5-6 : seam carving

including :

The result as following : reduce 50 pixel in horizontal direction

  • my implemention

  • opencv implemention

week 3-4 : fourier transform

including :

The result as following:

  • P1

  • P2

  • P3

 

week 1-2 : nonlinear filter

edit:

 * I took a mistake in original code(bilateral filter part) : I actually ran the opencv function when I wanted to test my function. So the time of calculation belongs to opencv, rather then mine. Thanks a lot for suggestion of professor Heidrich.

including :  

  • median filter (with opencv; with my implement)
  • min filter(naive; with PriorityQueue)
  • bilateral filter(with opencv; with my implement)
  • addtional work : guid filter

main source code click here

 

We compare different implement of those filter, and test their time. The result as following:

key point:

  • my implement of median filter cost much more time then opencv, if anyone has an idea about how to improve it, please contact with me.
  • when the kernel size of median filter < 70, it is faster to use quick sorting rather then histogram. because the average steps of 1 time calculate with histogram method ist 128, while the quick sorting is n*log(n), so when n is small, sorting method is the better choice.
  • my implemet of min filter with PriorityQueue is much slower then naive. Any idea about it?
  • my implemet of bilateral filter is a little bit faster then opencv, i also don't konw why.