Friday, March 06, 2009

making a facial model from a single photograph

A colleague sent me this youtube video, which I think is awesome:

They can construct a 3-d model of a human face from a single photograph, by combining two learning systems that look for facial features. See how authentic Tom Hanks's 3-d model came out:

I was surprised to see that this technology was presented in SIGGRAPH '99, an entire ten years ago, yet it's not prevalent today. The paper and other information can be found on this page about the Morphable Facial Model project.

Forget creating my own creature in Spore. I want to upload photographs of myself and everyone I know, and use them as the faces of the characters in the next game I play:


Performance might have been an obstacle in the past. The paper says each construction took 50 minutes on the SGI R10000 (approximately a 200MHz processor). The resulting model uses 70000 vertices. I bet that could be reduced, with some tradeoff in accuracy of the face.

I am waiting to use this technology with Rock Band. Also I think it would be popular among some people with the Hot Coffee mod.

15 comments:

  1. I remember seeing the video that year (a time when you had to download quicktimes because there was no Flash video...) and wow, I remember thinking that I was sure it would be commercialized "soon".

    It's my vague understanding that some X360/PS3 games (mostly sports games) already do face-mapping via their webcams, but I don't recall particularly awesome results.

    Have you seen the fidelity of hand-modelled faces (and textures) in today's top-end games? It's amazing.

    ReplyDelete
  2. Eons ago, I recall a group of guys building a prototype booth with a small camera system inside that would scan a paying customer, and output a Quake3 model file on CD. Wish I could remember the details of the camera setup (single camera / single pass around, or a four camera top-to-bottom scan, etc). The resulting models looked pretty good for a game in that era. :)

    ReplyDelete
  3. Yeah, and where's my jetpack? Weren't they flying people around in those at the '84 Olympics? That was 25 years ago!

    ReplyDelete
  4. Niniane, can you please add some links to the Siggraph paper you mention. Also from your other post - it would be great if you also write about software.

    ReplyDelete
  5. Application for 3d facial reconstruction is being used here:
    http://www.latimes.com/business/personalfinance/la-fi-himi8-2009feb08,0,1804104.story

    ReplyDelete
  6. This reminded me of a TED talk I recently saw : http://www.ted.com/talks/ed_ulbrich_shows_how_benjamin_button_got_his_face.html .
    They show how they created a model of Brad Pitt's face and used it in the recent film "The curious case of Benjamin Button". The objectives are very similar, although you can't just take a photograph and make 3D models.

    Also, the Stanford project http://make3d.stanford.edu/ comes to my mind. Given a 2D photograph, it comes up with a 3D model of it.

    ReplyDelete
  7. 你好啊,你应该是中国人吧,怎么你的日记都是英文呢~我在网上看到你的介绍,你离开GOOGLE的事在中国可真火~呵呵~我毕业一年半了,现在在LG做一个程序员,可以跟你交个朋友吗?(你肯定很忙吧...)

    ReplyDelete
  8. Sesh, I added a link to the project page which has the SIGGRAPH paper.

    ReplyDelete
  9. 3D model is cool,I hope I can make it someday.

    ReplyDelete
  10. this is really a surprise that a technology was available of such kind ..
    and that too ten years ago...
    well thanks for the information Niniane..

    ReplyDelete
  11. There was a similar program used to create the faces for Oblivion called "FaceGen." It provides morphology feedback in the form of anthropomorphic sliders. The software is not free but, last I checked, there was a free demo. It uses 2 photographs and a bit of user input to specify the location of major surface fiducials. I'm not sure how robust the algorithm is or if it even reproduces the head with any accuracy. My head came out a bit undersized and I've got a pretty large noggin'

    ReplyDelete
  12. Putting your face in a video game isn't so new, I think they did it with Tony Hawk Underground 1 or 2 and the Eye Toy. The same type of feature was also planned for Perfect Dark in junction with the Game Boy Camera but it didn't make it in the final game (pity, because for once I had both).

    The Hot Coffee mod... wasn't it that naughty trick in GTA?

    ReplyDelete
  13. I like how they turned the Mona Lisa's head, but preserved the cracks and brush strokes. It won't be long before we have advertisements generated in real-time, showing ourselves enjoying Product(tm), surrounded by attractive members of the appropriate sex (perhaps from our friends list on a social networking site).

    ReplyDelete
  14. I would buy something if I saw an ad of myself enjoying it along with my friends from a social network. That's such a good idea!

    ReplyDelete
  15. Check this out:

    http://gl.ict.usc.edu/Research/DigitalEmily/

    It's fast enough to do performance capture, and you get separate diffuse and specular. Me want.

    ReplyDelete

Your comment will need approval before it is shown: