How are trajectories organized and loaded in the dataset?

[shaojieb]

Hi!

Great work and I enjoyed reading the paper a lot.

While the paper/code is on FineGym, Diving48 and FisV, I wonder how in general I should process a sequential dataset (e.g., say I have a lot of facial mesh geometry frames) to fit the approach described in the paper. Specifically:

1. Do you extract trajectories of the same (temporal) length (e.g., a continuous 90 frames) after you extract the keypoints? (I assume so given the Table 2 in appendix.)
2. Do these trajectories overlap at all? For example, if you have a video v of 2000 frames at 30fps, are the trajectories something like v[:90], v[90:180], v[180:270], etc.? Or could there be a trajectory e.g., v[20:110], which overlaps with the v[:90] trajectory?
3. How are the timesteps t provided for samples in these trajectories? For example, will v[:90] have timesteps 1-90, and v[90:180] have (again) t=1-90?
4. Do you randomly select segments (e.g., past, future) from a trajectory? I'm particularly confused by this part as I did not find a clear answer in the paper. Given a long trajectory, how do you create segments from them, and how many do you create?

Thanks in advance for your answers. Again, excellent work!