some thoughts i had;
seems to me that you need an algorithm to first flag the outliers as possible noise, and confirm statistically with repetition. this is assuming that the noise is consistent enough that you get pixel repetition between at least a few frames.
in regards to the tilt-shift issue, i would think the solution would be to represent a 3d model from each frame and have the computer align the models, not to align the frames to create the model. what i mean by this is that you take your point grid in x-y-z space and draw lines between your points. hopefully the grid is triangular instead of square, but even if it's square, you cut the squares in half to create isosceles right-triangles instead of equal-lateral triangles. next, with your 3d triangular grid, you draw a surface between each set of 3 points. this gives you a rough model from each frame. i don't know what the frame rate of the connect is, but i would imagine that most 2-3 ghz processors should be able to do this calculation in the latency between frames - if not, the point grid could just be recorded in buffer and calculated as processor power becomes available.
using this technique, two 3d surfaces in virtual space are aligned according to a statistical best fit, thereby aligning the frames, and the camera tilt-shift can easily be calculated between frames. with this information known, we know that the change in velocity and rotation between frames is likely small, assuming a reasonably high frame-rate, so we can estimate the tilt-shift of the following frame, thereby making the lining up of future surfaces easier because the computer has a better starting point to calculate the best-match orientation and position of the surfaces.
as several surfaces become lined up, it becomes easy to statistically average the surface - not point - positions, disregarding the outliers, and creating an accurate 3d map.
another option which should not be discounted, would be to physically attach an accelerometer or a device with an accelerometer in it such as an i-touch, to the connect, thereby calculating the tilt-shift directly. this could be done in conjunction with the surface modeling calculation to provide better predictive tilt-shift calculations and model synchronization.
another thought i had, as people talk about doing all this in real-time, is that you don't really need to do it in real-time for a cave survey. all you need to do is record the data along with a time-stamp, then you can let a more powerful computer sitting at home, plugged into ac power do all the heavy lifting.
if the connect were employed along with a traditional point-to-point survey, and you could push a button when the connect were at each survey station to denote the station along with a time stamp, you would have both an accurate 3d model and an accurate set of coordinates to align the survey to. the connect is going to negate sketching, but i think point-to-point is still important in assuring the model's accurate alignment to 3d space. i think we still have yet to find the perfect point survey technique and excitement about the connect should not deter people from advancing point survey equipment and techniques.
another thought i had while thinking about accelerometers, is if something like the i-touch would be accurate enough to create a point-to-point survey just by carrying it from station to station. i doubt that it is, considering all the movements it would have to make in traveling such a distance in a cave and the error level that would create, but if it were that accurate... bottle of whiteout and an i-touch for a one-person point survey. combine that with one person scanning the walls, ceiling, and floor for the 3d model, and could have a 3d cave model aligned to an accurate real world 3d space about as fast as you could explore. imagine the possibilities... 2 miles of mapped cave on one trip.
hope some of those ideas help in your development. i'm really excited about what's going on with all this and happy to help in any way that i can. i'm not great with computer languages and thus poor at writing software, but i'm good with concepts and making things in my workshop, so let me know if i can be of any assistance.
peace, eli
p.s. i apologize for the lack of capital letters. i'm writing this on my girlfriend's computer and the shift key does not appear to be working so well right now.