Closed Mysterious-handsome-man closed 1 week ago
Hi, thank you
That's a simplifying assumption, because its hard to compute the rotation of each kernels in Spring-Mass system, so we assume that each Gaussian is a sphere of isotropic terms.
We only optimize a few parameters in registration network, the CUDA memory should be very close to static reconstruction. How many Gaussian kernels do you have?
Hi, thank you
- That's a simplifying assumption, because its hard to compute the rotation of each kernels in Spring-Mass system, so we assume that each Gaussian is a sphere of isotropic terms.
- We only optimize a few parameters in registration network, the CUDA memory should be very close to static reconstruction. How many Gaussian kernels do you have?
Thank you very much for your response. I have checked the input Gaussian models and indeed found the issue of excessive kernels. Once resolved, registration can proceed normally.
Additionally, I have another question: During the registration phase, it's necessary to assign an initial [R, T, S] to the 3D Gaussian model. How do you obtain this set of init_R, T, S? Is it manually set?
In our real world experiments, I use the first frame of each video for a coarse reconstruction just to obtain an approximate location of the object (which is the INIT_T I set). For INIT_R and INIT_S, just manually set.
Thank you for your excellent work, this is a very interesting project. And congratulations on your work being accepted by ECCV!! I have a few questions regarding debugging the code:
(1)I noticed that in the Gaussian_model_isotropic you replaced the original 3DGS's Gaussian_model use of 48-channel spherical harmonics for color control with a direct use of 3-channel RGB. What was the purpose of this change? Additionally, besides color, are there any other differences between the Gaussian_model_isotropic and the original Gaussian_model?
(2)I attempted to import the static Gaussian model generated by the original 3DGS, which includes color information across 48 channels in point_cloud.ply, into the registration network. However, during rendering, there was a massive increase in data volume, resulting in memory overload and rendering failure. Why does the registration network require so much data for rendering the static 3DGS model?