QuickTopic free message boards logo
Skip to Messages


Texture synthesis by non-parametric sampling

sameer agarwal
02:41 PM ET (US)
as a followup to this work, its worthwhile to look at the Image Analogies work done by Aaron Hertzman (Siggraph 2001). He takes the texture synthesis idea and carries it over to image synthesis by constructing analogies between sets of images. He is able to do

1. Texture synthesis
2. Texture Transfer
3. Painter style rendering

and what I find most intersting, texture by numbers
here is an example :

Dave KauchakPerson was signed in when posted
04:06 AM ET (US)
I think that's a great idea. In fact, there are a number of audio papers that address what might be considered texture for audio (including that one that we're going to look at in a couple of weeks). On top of "new-age" music, I it would be interesting to reproduce what I would consider "environmental" sounds, such as rain, the ocean, traffic, etc. I'm not sure if I'll have time to work it into the project for this class, but I will definately look into something like this in the near future.
Serge BelongiePerson was signed in when posted
02:26 AM ET (US)
Hsin-Hao: these both sound like great ideas! I especially like the ambient music synthesis idea...
Dave, maybe you can work this into your project?
Hsin-Hao Yu
12:52 AM ET (US)
Wow! What a cool idea. I am highly impressed. I have nothing substaintial to say about the algorithm, but I think some psychologists might find this algorithm very useful in the following situation: in order to find a brain area that is activated by viewing human faces (or objects, or anything), psychologists usually do a two block experiment, where in the first block you present the subject faces, while in the second block you use some non-face stimuli. The problem is that the non-face stimuli need to have very similar low-level strcture to faces (for example, average spatial frequency, color histogram..etc.) A simple way is to scramble the human faces. This however, introduces grid-like artifacts. It seems that to construct the non-face stimuli, you just treat the face images as textures and synthesis them. Can't wait to try it.

The other cool thing to do is to synthesis musical textures using this algorithm. Seems like a cheap way to generate some "new-age" ambiant music (as if they are not cheap enough already) or "sound-scape" a la Brian Eno and Robert Fripps.

Print | RSS Views: 577 (Unique: 397 ) / Subscribers: 0 | What's this?