Mockexam Question 6

Mockexam Question 6

by Loic Nathan Thome Busson -
Number of replies: 2

Hi,

I have a few questions for question 6:

- could you give an example of a CNN combined with a traditional algorithm for edge detection?

- What is "gradient-based tracking" exactly? I only see one slide about it (slide 69 on Edges lecture) but there are not much information about it (only that it "Maximize edge-strength along projection of the 3D wireframe.")

Thanks!


In reply to Loic Nathan Thome Busson

Re: Mockexam Question 6

by Benoît Guillard -

Hi Loïc,

- one naive example could be: along with the image from which you want to extract contours, you can provide your CNN with the output of a Sobel filter of your image. So the CNN receives (image , Sobel(image)), and does not have to "re-learn" Sobel, and can instead improve from this baseline. This is one very simplistic idea, but the main point here is: you don't have to throw all your traditional methods out of the window just because you are using CNNs.

- for gradient-based tracking, please have a look again at the lecture video: https://tube.switch.ch/videos/SRGPOSRMdo , at 1:06:45 . The idea is that once you have extracted edges, you can use them for a downstream application - here tracking. Is this clearer?


Best,

Benoît

In reply to Benoît Guillard

Re: Mockexam Question 6

by Loic Nathan Thome Busson -

Thanks for both your answers, this is really clear! For the gradient-based tracking, do we extract edges (meaning the output of Sobel filter or gaussian derivative filtering) or do we extract contours (meaning after applying Canny algorithm or equivalent CNNs)?

Loic