Raistlin said:
They could, but it would likely have unrealistic resource requirements.
For raw capture, like what's shown there, I think it's probably unlikely you'll see that sort of correction because it's just raw data before too much work is done to it besides creating the skeleton, but if you're animating an Avatar using that same data, you could employ a filter (and possibly some latency depending how aggressive it is) to prevent that sort of crazy point jumping from producing equally jumpy poses on the model based on a set of tolerance thresholds that maintain stability. I mean, a great deal of the basis for Kinect's pose estimation system is about working within what known limitations from constantly updated libraries about the human form and skeleton has, down to norms for how far a specific joint or bone can move.
Basically, I think the noticeable tradeoff, at this point, seems to be quality and stability of the end data to use for the game versus more latency introduced to perform more and better perform checks against larger libraries and enforce constraints. That's something they could keep refining and optimizing in a given situation, using known quantities. So, for instance, the more you use Kinect under your profile, the more data it could have stored on your specifics of range of motion of different extremities, understood metrics like your dimensions, and other data gathered such as known ideal (for that person for that profile) gestures. All of that could be shared and used by any Kinect title to better suit itself to you and possibly eliminating the future need for specific calls for a calibration. A constantly evolving profile of limits, ideals, etc. I believe that's how Kinect already works, at least on some level. And along with a more refined and a better understanding of you, from voice to movement to appearance, better optimization for traversing the trees of possible poses is custom tailored to you...possibly resulting in less latency and, most certainly, an output in games that is far less prone to error. Basically, a problem of heuristics.
Eventually, on better hardware with more and faster memory and better quality of capture, you could have AIs that could even learn your tendencies in a way that has long been limited to very specific and known range of possibilities, like estimating how you'll move on the chessboard, but in a case where the range is a great deal larger, with more spatial possibility, like trying to intercept an attack coming from your body before you actually have committed to the movement enough for it to determine with greater accuracy due to the need to act now because of the speed of attack involved. Real learning mechanisms with important local histories/memories that can be accessed by individual AI entities or shared in any fashion with many or groups or just those that can 'witness' it and even recognize it with training.
For Kinect on X360, though, I think it'll get faster and more efficient based on application specific optimization. So, by second gen releases as late as end of next year to mid-to-late 2012, the software should have made marked improvements in speed, accuracy, and general range of use. I think the most incredible uses will end up being more believable and 'alive' virtual characters in more densely useful and interactive virtual environments with a focus on virtual props and tools that have more simulation elements and, in general, more complicated mechanical workings for direct interaction during action game moments. The direct body to body mapping stuff should obviously improve, but I think a strong blend of gestural use and 1:1 mapping will be implemented in most games to ensure the most fun and usable mix abstraction of game activity and player activity. I know a lot has been made about 1:1 being this holy grail of motion control or whatever, but games are still mostly about learning a language of game mechanics and gestures can work better and be more efficient than having super-fidelity movement as the main input. Considering fatigue and being friendly to the player through longer play sessions is a requirement if you're going to do more than short-burst-style games.
)
at the flipping out skeleton
