The training speed of 4DGS-SLAM

[MYM181]

Hi, thanks for your great work!
I run the code successfully. It seems that the training speed is relatively slow on a single A100(40G) GPU.

I was wondering if you could provide some information on its expected performance. Specifically:

• What's the normal FPS for the mapping algorithm on your end?
• Where does the majority of the time cost come from?

Any insights would be super helpful. Thanks again for your time!

[vamWu]

Thank you for the Author's valuable contribution and for making the project open-source.

1. I have also reproduced the project and noticed the same issue. Based on my understanding, the computation of optical flow is quite time-consuming. I would like to inquire how the runtime speed can be improved to achieve an online performance similar to the original MonoGS.

2. In addition, I noticed that in the code, dynamic_render is set to False and that many lines related to dynamic_render have been commented out.
   Could you please clarify the roles of dynamic_render and dynamic_network? My current understanding is that "dynamic_render: False" indicates that the current rendering process is static, while dynamic_network refers to training a dynamic deformation network.

[yanyan-li]

Hi guys, thanks for your interest and glad to hear you got the code running successfully!

Yes, the training speed is indeed still an open challenge in this domain. The main bottleneck comes from training the transformation field online, which is computationally expensive. This step is essential for handling dynamic scenes, but it does introduce noticeable overhead compared to purely static mapping approaches. The following performance analysis tested on the BONN sync sequence.

Modules	RoDyn-SLAM	SplaTAM	MonoGS	DG-SLAM	Ours	Modules	SC-GS	Ours
Static Tracking						Dynamic Mapping
Track. Iter. [ms]	15.7	6.5	10.0	6.5	5.6	Map. Iter. [s]	0.3	0.5
Track. Frame [ms]	314	1300	82	54	283	Total Time [h]	4.9	2.8