My DSO fail story

First, the way to use DSO using ROS, I made a tutorial

• https://youtu.be/YPkSXTaGACA

I used a FLIR’s A65 thermal camera and the details are given in this paper

• Sparse Depth Enhanced Direct Thermal-infrared SLAM Beyond the Visible Spectrum (https://irap.kaist.ac.kr/publications/yshin-2019-ral.pdf)

I used the intrinsic calib like:

• RadTan 4.498560000000e+02 4.508560000000e+02 3.291928000000e+02 2.502157000000e+02 -0.406290 0.125974494 0.001216395 0.001754115
  640 512
  crop
  640 512
  0

and play the node

• $ rosrun dso_ros dso_live image:=/thermal/image_raw calib=./camera_thermal.txt

Results:

• sequence 1: system stopped…

• sequence 2: scale so shrank.

• sequence 3: the system flyes to the sky after the major occlusion by a human

Difficulties I think:

• I used the unnormal camera (i.e., thermal camera); but the intrinsic calibration is well done.
• The camera is installed perpendicular to the moving direction (i.e., left-side)
• The current sota geometric-based SLAM still suffers and shows low performance in a highly dynamic environment.

Questions

• DSO does not have re-localization? (system just stopped and cannot survive again)
• DSO does not support sim3 optimization? (scale inconsistency easily occurs than ORB-SLAM2)