sergeyprokudin/dpf

Dynamic Point Fields: Towards Efficient and Scalable Dynamic Surface Representations (ICCV 2023)

dpf's Issues

Projection Matrix

Thanks for your nice work! I noticed Pytorch3d's FoVOrthographicCameras is used in your code. What makes me wonder is why orthogonal projection is used. Will it have any impact on performance if it is changed to perspective projection?

Questions about Static surface reconstruction with point sets

Hi, thanks for your nice work! Table 1 of the paper shows the powerful overfitting ability of point cloud for static surface reconstruction. What confused me is the comparison of ngp baseline. Because it seems that static surface point cloud optimization method is initialize with the point cloud sampled from gt mesh, and ngp and other implict methods cannot do this. Can you tell me what the losses of ngp optimization used, and how to extract point clouds from optimized npg to evaluate? Thanks.

Question about the "Single Scan Animation" Code

Hello, thanks for the great work!

I have some questions about "Single Scan Animation" demo code.

In Single Scan Animation demo, I realized that you only used vertice guidance loss and didn't use chamfer distance loss to train the deformation model. Why do we not use chamfer distance loss in this case?

In the demo, we only train our siren mlp to a single pair of deformation (S_A to S_B) and generalize to unseen poses. What if we can train multiple pairs of deformation (S_i to S_j) and generalize to unseen poses? Does this help improve novel pose results? If we train our siren mlp with only a single pair of deformation (S_A to S_B), is this working with another shape pair (S_C to S_D) of the same subject?

Chamfer Distance Loss

Hello,
thanks for sharing the great work..!
When I look at the code for colab surface reconstruction,
it seems that the normal discrepancy mentioned in the paper is not used in the chamfer distance loss.
I'm curious about the reason for this.
Additionally, did you experimentally perform normal estimation using Poisson solvers as mentioned in the appendix?

Ignore part of the target shape

Hi,
Congratulations on this, I am getting really impressive results when matching/fitting parametric template heads to scanned heads. One question, is it possible to fit to the closest area and ignore the rest of the target shape (when both are already quite close in shape)? Now I am trying to fit a template head to a scan that has the head and the whole body, so it tries to fit the head to the whole body.
Thanks!
Daniel