Open dlazares opened 7 months ago
Hi@dlazares, thank you for your interest. The temporal gradient demonstrates its benefit only when the relevant threshold is appropriately configured. However, it is laborious to tune these hyperparameters, and the optimal value varies across different scenes. Therefore, in the open-source code, to keep the config as consistent as possible for different scenes, we only retain the sampling in 4D distribution, which is proved more important than considering the time gradient as a criterion for densification and is already enough to achieve sufficient performance.
In your paper you describe "In terms of density control, simply considering the average magnitude of view space position gradients is insufficient to assess under-reconstruction and overreconstruction over time. To address this, we incorporate the average gradients of µt as an additional density control indicator".
In the code, I see the variables passed through to the densify function but they're not used. Did you ultimately find these not helpful? could you explain why they were or were not helpful in the end?
Thanks for open sourcing your work! It's a great paper, love the 4d representation.