Mock Exam (2020) - Q9

Mock Exam (2020) - Q9

by Yasamin Borhani -
Number of replies: 5

Hello,

I was wondering if the answer to Q9 of the mock exam (2020) is correct. As far as I understood, the visual hull algorithm can't correct the silhouette image, it's able to use ill silhouette images to reconstruct the 3D shape but the purpose of that I guess is not to fix the silhouette images, and it doesn't directly improve that. Could you please tell me if I understood correctly or not?

 

Thanks in advance.

Best regards,

Yasamin


In reply to Yasamin Borhani

Re: Mock Exam (2020) - Q9

by Deniz Sayin Mercadier -
Hello Yasamin,

While it's not the only purpose of the algorithm as you say, the visual hull can definitely be used to refine a silhouette. Imagine an image A where the object of interest is occluded by some other large object. The silhouette/mask of the object will be partial in this case. But if you can reconstruct a visual hull from other images, then it would be possible to project the hull back into image A to complete the missing part of the silhouette, despite it being occluded in the original image. While it will not be as precise as actually having the full object on camera, the approximation should be pretty good, depending on the shape of the object and positions of the other cameras.

While they should be pretty clear, let me also go through the other choices for the sake of completeness:
2. The visual hull does not capture texture, it uses no color information, only binary masks.
3. The visual hull can only capture a convex hull of the object, and will not be able to capture the concavities.
4. Each new image/camera will further constrain the convex hull and possibly make it *smaller*, closer to the original object's smallest possible convex hull. Making it bigger would be worse.

Best.
In reply to Deniz Sayin Mercadier

Re: Mock Exam (2020) - Q9

by Ahmad Bilal Kakar -

Hello, 


I am a bit confused about point 4 in you answer. Why does the convex hull become smaller? Isn’t the case when the camera is not calibrated? 


Thanks

In reply to Ahmad Bilal Kakar

Re: Mock Exam (2020) - Q9

by Deniz Sayin Mercadier -
Hello Ahmad,

The visual hull is entirely dependent on the cameras' view cones. If you add new cameras that can reach other cameras' "blind spots", you can have a better convex hull. I'm sure these sentences don't make much sense, so here's a crude visualization in 2D:
visual hull change
(image from https://homepages.inf.ed.ac.uk/rbf/CVonline/LOCAL_COPIES/AV0809/schneider.pdf)
The thick green outline is the hull with cameras C1..C4. Notice how some parts of the hull protrude a lot from the original shape due to the lack of information around those spots. If we add another camera C5, it will cover part of these blind spots and make the reconstructed hull smaller. Essentially, each new camera shaves some parts off the existing hull. The more cameras the better, and with infinite cameras having perfect coverage we would have the smallest possible convex hull. Hope that makes sense!

Best.