Are GPU reconstructions less precise than CPU ones

WebODM
#1

I have uploaded the same dataset to two different webodm servers. One with lots (128) of cpu cores and other with less cores (16), but with a good gpu (A100) to see which one will run faster. On both systems, reconstructions settings were set to 3D model, which is my goal.

GPU one took 1:20, CPU one took almost twice as long at 2:40. Clearly GPU reconstruction is winner here. However, visually CPU derived one looks better aligned as a 3D model than the GPU one (below are the links for both datasets). I haven’t done any quantification yet, but would like to hear people thoughts, if anyone has encountered something like this, or have suggestions?

GPU derived 3D model
CPU only 3D model

2 Likes
#2

It seems like CPU feature extraction and detection might be more robust and lead to a better sparse/dense cloud, so you may have to crank some parameters up on GPU to match that. There should still be a time benefit, I’d think.

1 Like