bouteillebleu (![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png)
bouteillebleu) wrote2007-11-28 02:09 pm
bouteillebleu) wrote2007-11-28 02:09 pm
Cambridge University SciSoc talk yesterday - "Computer Science and Illusion"
I went to a talk yesterday on Computers and illusion - from photography to colour vision. it was interesting, though due to lack of time he cut the part about shading and colour vision very short, which was a pity as I'm interested in colour vision.
The talk
The talk covered four aspects of vision, and started each section with examples of visual illusions, then explained the theory of vision that made the illusions work, then an implementation of that theory for computer vision.
Sections were perspective, texture, shading and stereoscopy (stereo vision).
Perspective
Possibility of interpreting line drawings by labelling their edges, and constraining which edges can be convex, concave, occluding or normal by what types of edges can meet at junctions. I think he cited David Waltz here (the PDFs on Waltz's website unfortunately do not work, but they give the titles of articles he has written).
There was further work done on this by Sugihara in his 1986 book Machine Interpretation of Line Drawings (full text available in PDF form).
Texture
Some more stuff here, with a lot of hints about how commercial image processing programs manage to do clever things with filters. I didn't take many notes other than a few words of what each paper was about, but there are some quite cool things (such as Freeman, "the man with the face made of rice").
The idea of a "Markov Random Field" was brought up. Something like a Markov Chain but in two dimensions (or more?). As one of the other CSTIT students I met there commented, Markov got his name on pretty much anything probabilistic. :)
Some more specific papers were mentioned:
* On texture synthesis, work by Efros and Leung and De Bonet and Viola. Efros and Leung have a very neat example of texture synthesis from a picture of a block of text, which when the size of each texture patch is less than the size of a letter, produces some interesting novel glyphs that aren't in the Latin alphabet.
* On style transfer, also known as "image analogy" or "how to make a photo look like an oil painting", work by Salesin, Oliver et al.
* On texture transfer by quilting, also referred to above as "the man with the face full of rice", work by Freeman and Efros. The "face full of rice" is an example in the paper of taking an image of Freeman's face, then replacing the textures in the image with those of a bowl of cooked rice, thus producing a picture of a bowl of cooked rice with a face visible.
* The lecturer showed some of the work that had been done at Microsoft Research on a project called PatchWorks, which included ways to use texture synthesis to remove objects from images almost seamlessly.
* On dynamic textures, for example "how to turn a short film of a flame burning into a much longer one by synthesising more of the flame's movement", work by Soatto, Doretto and Wu and by Fitzgibbon.
* On Bayesian matting, for example "how to take a picture of someone with very fine hair, change the background they're standing in front of, and still keep all the hair looking right", work by Chuang, Curless, Salesin and Szeliski.
Shading
This bit was unfortunately cut because the lecturer was running out of time. I'll have a look around and see what's been done in this field, but I didn't get any citations from the lecture itself.
Stereoscopy
* On using multiple cameras to get the advantages of human depth perception and other aspects of humans-having-two-eyes, work by Kolmogorov, Criminisi et al. [link goes directly to a PDF].
One of the questions after the talk was whether work has been done with more than two cameras - apparently, three cameras gets you much better results than two, giving computers something of an edge in depth perception that humans do not have.
The talk was given by Andrew Blake of Microsoft Research Cambridge. MSR's page of articles on computer vision is also a good read.
The talk
The talk covered four aspects of vision, and started each section with examples of visual illusions, then explained the theory of vision that made the illusions work, then an implementation of that theory for computer vision.
Sections were perspective, texture, shading and stereoscopy (stereo vision).
Perspective
Possibility of interpreting line drawings by labelling their edges, and constraining which edges can be convex, concave, occluding or normal by what types of edges can meet at junctions. I think he cited David Waltz here (the PDFs on Waltz's website unfortunately do not work, but they give the titles of articles he has written).
There was further work done on this by Sugihara in his 1986 book Machine Interpretation of Line Drawings (full text available in PDF form).
Texture
Some more stuff here, with a lot of hints about how commercial image processing programs manage to do clever things with filters. I didn't take many notes other than a few words of what each paper was about, but there are some quite cool things (such as Freeman, "the man with the face made of rice").
The idea of a "Markov Random Field" was brought up. Something like a Markov Chain but in two dimensions (or more?). As one of the other CSTIT students I met there commented, Markov got his name on pretty much anything probabilistic. :)
Some more specific papers were mentioned:
* On texture synthesis, work by Efros and Leung and De Bonet and Viola. Efros and Leung have a very neat example of texture synthesis from a picture of a block of text, which when the size of each texture patch is less than the size of a letter, produces some interesting novel glyphs that aren't in the Latin alphabet.
* On style transfer, also known as "image analogy" or "how to make a photo look like an oil painting", work by Salesin, Oliver et al.
* On texture transfer by quilting, also referred to above as "the man with the face full of rice", work by Freeman and Efros. The "face full of rice" is an example in the paper of taking an image of Freeman's face, then replacing the textures in the image with those of a bowl of cooked rice, thus producing a picture of a bowl of cooked rice with a face visible.
* The lecturer showed some of the work that had been done at Microsoft Research on a project called PatchWorks, which included ways to use texture synthesis to remove objects from images almost seamlessly.
* On dynamic textures, for example "how to turn a short film of a flame burning into a much longer one by synthesising more of the flame's movement", work by Soatto, Doretto and Wu and by Fitzgibbon.
* On Bayesian matting, for example "how to take a picture of someone with very fine hair, change the background they're standing in front of, and still keep all the hair looking right", work by Chuang, Curless, Salesin and Szeliski.
Shading
This bit was unfortunately cut because the lecturer was running out of time. I'll have a look around and see what's been done in this field, but I didn't get any citations from the lecture itself.
Stereoscopy
* On using multiple cameras to get the advantages of human depth perception and other aspects of humans-having-two-eyes, work by Kolmogorov, Criminisi et al. [link goes directly to a PDF].
One of the questions after the talk was whether work has been done with more than two cameras - apparently, three cameras gets you much better results than two, giving computers something of an edge in depth perception that humans do not have.
The talk was given by Andrew Blake of Microsoft Research Cambridge. MSR's page of articles on computer vision is also a good read.