Speech recognition software performs best when you train the software by speaking a series of important words and sentences. It seems to me that the technology already exists to make prototypes of a system that would embed videos with full text transcriptions of everything spoken on camera as metadata. Nevermind the fact that if what you were shooting had a screenplay you could feed the entire screenplay to the camera before you shoot so it can anticipate what the actors will say.
Techniques such as dynamic time warping would accomodate the variations between different takes of the same shot so that they are all regarded as versions of one another. Seems like mpeg7 would be the best possible way to store this data today.
Frameline 47 lets you take advantage of the mpeg7 standard quite fully and its not surprising at all that a company like Eptascape would be in the security and surveillance market. What we need is the metadata to be entered half-automatically during the shooting process combined with a learning curve for the camera where you have it learn what the actors look like and sound like. A sophisticated infra red camera would help define actors from the background. Using three cameras (one on the X, Y and Z axis) could help to define the physical space and depth. And now a dog in infrared:
Posts
No comments:
Post a Comment