3 ms·

What about the source data makes it so difficult to clean up? It feels like nothing has progressed since the days of the first ping pong ball suit even though

What about the source data makes it so difficult to clean up? It feels like nothing has progressed since the days of the first ping pong ball suit even though cameras and image processing have become infinitely more capable.

baddox 8y ago
I can only speculate. For faces, I'm sure there's tons of manual work to be done to reduce the uncanny valley, especially when it's being mapped onto a different facial geometry (like King Kong). For body movements, perhaps the collisions and interactions with the virtual world require a lot of manual tweaking to get them looking realistic.
blihp 8y ago
The fact that there's so much of it and the sources are indirect and imprecise. All motion capture with actors is from the surface of their body and then you have to reverse-engineering all that data into skeletal and facial muscle positions.[1] The techniques are very cool but not perfect and there is error at every step of the process.[2] So then you still need tools and people go back through the data to clean it up / smooth it out... and even then, it's still not going to be perfect. [1] For some strange reason surgical procedures to implant hundreds of hyper-accurate mo-cap sensors in the actors or otherwise hack into their nervous system hasn't become a thing. ;-) [2] Camera lenses aren't perfect, dot placement (for marker based systems) isn't exact, tracking software slips since the markers aren't infinitely small and can move around a bit etc.
- danmaz74 8y ago
  This looks like an area where machine learning could really help... Have you got any idea if there is progress here?
  - blihp 8y ago
    I agree that it's a likely application which would reduce the manual effort involved. Someone is probably doing something with it (is there anything someone isn't trying to apply machine learning to these days?), but I'm not familiar with anything specific.
    - jzl 8y ago
      https://schedule.gdconf.com/session/a-new-era-of-performance-capture-with-machine-learning/860568 https://schedule.gdconf.com/session/a-new-era-of-performance...

baddox 8y ago

I can only speculate. For faces, I'm sure there's tons of manual work to be done to reduce the uncanny valley, especially when it's being mapped onto a different facial geometry (like King Kong). For body movements, perhaps the collisions and interactions with the virtual world require a lot of manual tweaking to get them looking realistic.

blihp 8y ago

The fact that there's so much of it and the sources are indirect and imprecise. All motion capture with actors is from the surface of their body and then you have to reverse-engineering all that data into skeletal and facial muscle positions.[1] The techniques are very cool but not perfect and there is error at every step of the process.[2] So then you still need tools and people go back through the data to clean it up / smooth it out... and even then, it's still not going to be perfect. [1] For some strange reason surgical procedures to implant hundreds of hyper-accurate mo-cap sensors in the actors or otherwise hack into their nervous system hasn't become a thing. ;-) [2] Camera lenses aren't perfect, dot placement (for marker based systems) isn't exact, tracking software slips since the markers aren't infinitely small and can move around a bit etc.

danmaz74 8y ago
This looks like an area where machine learning could really help... Have you got any idea if there is progress here?
- blihp 8y ago
  I agree that it's a likely application which would reduce the manual effort involved. Someone is probably doing something with it (is there anything someone isn't trying to apply machine learning to these days?), but I'm not familiar with anything specific.
  - jzl 8y ago
    https://schedule.gdconf.com/session/a-new-era-of-performance-capture-with-machine-learning/860568 https://schedule.gdconf.com/session/a-new-era-of-performance...