About the training

[liziwennba]

Hi, thank you for your excellent work! I'm curious about your implementation. It seems that your code of scene velocity is not the same as you mentioned in your paper. And you scaling t is implemented by activation equation torch.exp. Can you provide the reason behind this?

[Fumore]

Hi, scene velocity only an estimator of point's motion, and we find that it is more effective to define it this way. I think that any scene velocity defination that is inversely proportional to $\rho$, equal to the instantaneous velocity at $\tau$ when $\rho=0$ and equal to 0 when $\rho \to \infty$ is appropriate. And in terms of opacity decay, i.e., scaling t, we just used the gaussian family in as the 3DGS do for scaling xyz. We will update our paper in arxiv.