search

Colmar-zlicheng / Spring-Gaus

[ECCV 2024] Reconstruction and Simulation of Elastic Objects with Spring-Mass 3D Gaussians
https://zlicheng.com/spring_gaus/
Apache License 2.0
56 stars 0 forks source link

Some questions about the paper #2

Closed guanjunwu closed 3 months ago

guanjunwu commented 4 months ago

Hi, thanks for your great work! I have some little questions about the paper:

  1. It seems like anchor points $A$ are attached to initial 3D Gaussians?
  2. How long does spring-gaus is trained on the real-world samples? It looks like the paper only present details on synthetic data.
  3. what is registration network? As is shown in the paper, three viewpoints are insufficient for 3D Gaussians, it is related to sparse-view reconstruction and I think it also can be a interesting and difficult problem. How can you train the dynamic 3DGS/4DGS on the sparse-view input, can you provide more details?

Yours sincerely, Guanjun

Colmar-zlicheng commented 4 months ago

Hi Guanjun, thank you for your issue.

  1. Anchor points are independent from 3D Gaussians, we just compute anchor points' initial position from Gaussian kernels' centers.
  2. Almost same to synthetic data (Static reconstruction will be longer due to the bigger resolution of images. But in dynamic stage we use the same number of anchor points, so the training time will be similar.).
  3. We first use 50-100 images to train a static 3DGS, which has different position/rotation and shape with the initial state of 4DGS. So, we use a registration network to align static 3DGS and initial 4DGS.

Best wishes, Licheng