Low speed on TUM/Replica dataset

[Poiw]

Hi,

I tried to run the pipeline on TUM and Replica dataset but it is much slower than the numbers reported in the paper. The actual speed is less than 0.5 FPS. I tested the pipeline on an RTX 4090, so it should be faster. Is there any settings I need to modify to match the numbers in the paper?

Thanks

[eriksandstroem]

Hi @Poiw ,
Sorry for my late reply. The reason for this traces itself back to MonoGS actually, since we reuse parts of their mapping code. It seems that running the codebase on a cluster machine is much slower than running it on a local device, as long as the local device has enough memory etc. We are still not sure why this is the case.

Did you run on a local machine or on a cluster?

[Poiw]

HI @eriksandstroem ,

Thanks for the reply! I run the code on a local machine. The replica sequences are faster which achieves > 1fps, but the TUM sequence (fr1, desk) I tested is less than 0.5 FPS.