D120.de/forum

Hi,

i don't have the code in front of me right now, but im2double() automatically scales the image to [0,1]. Try to typecast it with double(x) instead. Then, it should be double in the range of [0,255]. If that does not fix it, let me know and i will investigate further.

Regards,
Thorsten

Hi, i wrote an reply, but apparently i forgot to press send and now it is lost. :roll: Anyways, lustiz's reply pretty much answers your question. Of course it makes only sense to do NMS for each pixel. Non overlapping windows would not be able to suppress surrounding values consistently. Think about...

Hi, 1) Does it make sense to do "standard" convolution here? Elementwise matrix operations on differently sized matrices are going to be difficult, right? For boundary handling "symmetric" is a good choice, but in principle they all have different pros/cons. 2) You can set the range yourself by imsh...

Hi, i'm soory i obviously didn't get where you were coming from exactly. So yes you could solve this with a vector (-1,0,1) for a filter with size 3. However you are note quite right with your solution for filter size 4. This is obviously not equally spaced. Think about a grid of pixels (5x5). Where...

Hi, i'm not really sure why you write the first derivative (central difference) there. You are supposed to create a isotropic gaussian as depicted in lecture 3 slide 17. Think about the goal that we try to achieve here and what the actual purpose for convolution with a gaussian filter is. I hope thi...

The first assignment will be out soon. All material will be available in Moodle and on the webpage. Additionally, we will announce these matters over the mailing list.

Kind regards
Thorsten

1. The order is not of relevance, since you build a sparse matrix (sparse()) out of this list. This is basically an intermediate step to building the "pairwise" variable. 2. You basically create an adjacency matrix, where only the upper triangle is used. The GCO library assumes undirected edges as s...

Hi,

your solution is correct. it contains all the edges of a (3,2) image in a list.

You should implement this without any loops. Loops are only allowed when implementing iterative procedures like iterating over labels in alphaexpansion(...).

Regards,
Thorsten

Thanks for your hint. Assignment 3 is now visible in Moodle.

Regards,
Thorsten

Robert is right regarding the definition. The formal definitions from both books ("Probabilistic Graphical Models: Principles and Techniques", "Computer vision:models, learning and inference") suggest that the markov blanket of variable X does not include X itself. However, it is convention (c.f. "C...

Did you already answer 3.3? What is the reason for computing the log instead of working with the actual probability densities? Does the negative log-likelihood give you a reasonable results? Think about what numbers are actually multiplied, when you compute the gaussian likelihood. Once you look at ...

The joint distribution is given as p(v, w, x, y, z). It has 5 variables. Now you should marginalize p(v, w, x, y, z) with respect to w and z. After that you should marginalize over x. You should interpret the first part of the first sentence as: "You are given a distribution p(v, w, x, y, z) over 5 ...

You are absolutely right with your interpretation.

Regards,
Thorsten

Exercise2 is done and can be picked up at Carola Eichel's Office.
The corresponding points will be entered into webreg in the next hour.

Exercise 2 is now available on the course website. This exercise is due before June 30, 2pm.