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
73 stars 0 forks source link

The precise geometric structure of the Gaussians in the Static phase #7

Open Mysterious-handsome-man opened 4 weeks ago

Mysterious-handsome-man commented 4 weeks ago

In the Static phase of your project, an initial Gaussian is trained. I reviewed the point clouds of these Gaussians (for example, the hamburger), and I found their geometric structures to be remarkably precise, without any noise or artifacts. How did you achieve this? My initial guess is that this is due to the fact that all camera poses in the training data are strictly calibrated, which prevents the Gaussians from developing any artifacts or ellipsoids outside their geometric structures. Do you think this hypothesis is correct? Or did you incorporate other operations to eliminate artifacts?

Mysterious-handsome-man commented 4 weeks ago

In the Static phase of your project, an initial Gaussian is trained. I reviewed the point clouds of these Gaussians (for example, the hamburger), and I found their geometric structures to be remarkably precise, without any noise or artifacts. How did you achieve this? My initial guess is that this is due to the fact that all camera poses in the training data are strictly calibrated, which prevents the Gaussians from developing any artifacts or ellipsoids outside their geometric structures. Do you think this hypothesis is correct? Or did you incorporate other operations to eliminate artifacts?

image

Colmar-zlicheng commented 3 weeks ago

Hi, for static reconstruction of real object, we have lots of camera views calibrated by colmap.