Repeatability of Experimental Results on ICL Datasets

[zj2013]

Thank you very much for your open source code, which has greatly inspired me. When I use the default setting file dataset_ Icl.yaml and reslam_ Settings.yaml tests the two sequences lr-kt0 and lr-kt1 of ICL, and the experimental results are as follows: ATE of VO+LM lr-kt0: 4.5cm lr-kt1:1.8cm

ATE of VO+LM+LC lr-kt0: 7.5cm lr-kt1:1.8cm

However, the experimental results in your paper are ATE of VO+LM lr-kt0: 3.2cm lr-kt1:1.8cm

ATE of VO+LM+LC lr-kt0: 1.9cm lr-kt1:1.7cm

Do I need to change some parameters of the system to achieve the same results as in the paper?

[fabianschenk]

Hi @zj2013 ,

I did the evaluation almost 4,5-5 years ago and can't remember what parameters I used and if it was exactly the same as in the settings.

One thing to keep in mind is that the system is not not fully deterministic, so multiple runs will lead to typically slightly different results:

• The random ferns used for LC candidate selection are based on randomly sampled feature points
• Numerical differences due to multi-threading and float operations