Question about the Scale Handling

[AIBluefisher]

Hi @wbhu

I find that the paper uses an interesting approach to handle scale across different datasets, using the log of the momentum and a scale-aware gating mechanism. I wonder if you tried normalizing the camera poses so we can simply encode the plucker encoding E([d, m])?

[mhyqwe]

Yes, we did try this before. In one ablation we normalized every clip's trajectory to unit scale and fed the raw 6D Plücker E([d, m]) directly, without the log‖m‖ decomposition or the scale gate.

It trains fine, but it underperforms on pose following / camera-pose eval. The reason is that normalizing each trajectory to a fixed scale throws away the absolute scale, and that scale signal is actually useful — especially once you mix metric and up-to-scale pose sources in one model. We plan to include this ablation in a follow-up.