How to adjust parameters to trigger "Global loop closure"?

[HTLife]

Currently I'm working on applying TUM "fr2/360_kidnap" to ElasticFusion. However, loop could not been detected. (Kintinuous could detect loop but the map is broken.)

If I want to get a map with complete reconstruction result and loop closure, what would you guys suggest?

Here are my experinment parameters and result

ElasticFusion

./ElasticFusion -l '/dataset/tum/rgbd_dataset_freiburg2_360_kidnap/scale_0_0002.klg' -ftf -d 12 -c 4 -ft 0.1

• Do frame-to-frame RGB tracking.
• Cutoff distance = 12m
• Surfel confidence threshold = 4
• Fern encoding threshold = 0.1

Final result screenshot: Video: https://youtu.be/PWe2Cziccw8

I provide my .klg file here for result reproduce. (ftp://ccurvl.csie.io/slam_exp/)

I did both ElasticFusion and Kintinuous on TUM "fr2/360_kidnap"

• ElasticFusion: [good] map reconstruction, [bad] loop detection
• Kintinuous: [good] loop detection, [bad] map reconstruction (maybe didn't find the right voxel threshold or other related...)

I also tried the instruction in https://github.com/mp3guy/ElasticFusion/issues/21 to adjust some parameters, but loop detection still don't work with TUM kidnap dataset.

---

Kintinuous

Seems like depth cutoff issue, still finding related code or fix this problem. ./build/Kintinuous -v ./vocab.yml.gz -l /rgbd_dataset_freiburg2_360_kidnap/scale_0_0002.klg -ri -fl -od -dg 0.1 -cw 20

[mp3guy]

Please see Page 16 Table 2; http://www.thomaswhelan.ie/Whelan16ijrr.pdf

You would need to write new code to manage relocalisation and map merging if you want either of these methods to work on kidnapped sequences.

[HTLife]

@mp3guy Thank you for your quick response. Before writing new code, I want to test the capability of place recognition module (both DBoW2 and Fern). After trying out some fern related setting, loop still couldn't been correctly detected (ex. Fern encoding threshold = 0.1). Would that be the limitation of "fern" method? Or I didn't tune it right?