search

graphdeco-inria / reduced-3dgs

The code for the paper "Reducing the Memory Footprint of 3D Gaussian Splatting"
Other
96 stars 6 forks source link

if you have tested the changes in the use of VRAM for graphics card rendering before and after compression? #3

Open smart4654154 opened 3 months ago

smart4654154 commented 3 months ago

Thank you very much for your work The data compression work is excellent May I ask if you have tested the changes in the use of VRAM for graphics card rendering before and after compression?

smart4654154 commented 3 months ago

What I mean is to use specific data testing to obtain changes in the usage of VRAM

smart4654154 commented 3 months ago

in the parper: We note that integrating codebooks and half-float type handling into the renderer would enable the same reduction for a scene’s VRAM consumption as for its required disk storage.

May I ask if you have specific test results?I want to know the specific changes.thank you

PanagiotisP commented 3 months ago

Hello and thanks a lot for the question. The short answer is no, we don't have specific test results for VRAM consumption reduction.

The long answer is that runtime VRAM requirements are difficult to compute from the primitives' count and details. In our paper, we focused on representation size, that is stored in disk file size which is why we don't report any numbers about VRAM consumption during training or during inference/testing. Hence, no specific test results have been acquired. However, we note that the decrease in point numbers directly translates to less VRAM during rendering (proportional but not the same % as the runtime has some memory overhead) and for the other two approaches a more involved implementation can be made to reduce the runtime memory requirements (a quick and dirty implementation for the variable SH bands is included in the code). The amount of the actual and potential reduction though is difficult to calculate. Hope that helps