Improve Hand Poses Stability

[andrearosasco]

I'm kind of playing ahead with this one since, until we actually start closing the loop, I don't know if the noise of the hand poses is too much.

Anyway, the hands were jiggering a lot because of the variability of:

• partial point cloud
• subsampling
• denoising
• plane fitting

I tested different solutions and the best one I've found is to tackle the problem upstream. The solution consists of fitting planes on the input point cloud.

1. We fit two planes to the current partial point cloud (if we can't, just a face is visible hence we skip the reconstruction)
2. We measure the fitness score of the previous iteration planes on the current point cloud
3. If current_plane_score - previous_plane_score is within a certain threshold we keep using the previous output; otherwise, we recompute everything.

[andrearosasco]

With the right threshold, we keep the old hand poses if the box doesn't move too much

And still, be able to track the box when the point of view changes