How to adjust parameters for Insta360 ONE X2?

[Shunmo17]

I use tracking and mapping of OpenVSLAM with RICOH Theta V and Insra360 ONE X2. Both cameras' config files were build based on the example file for RICOH Theta S 960 (https://github.com/OpenVSLAM-Community/openvslam/blob/main/example/aist/equirectangular.yaml), and I only changed video resolution and max_num_keypoints.

In order to evaluate the accuracy of OpenVSLAM, I ran it with the same video several times and compared these two cameras, I noticed that the estimated trajectories in each trial of Insta360 ONE X2 vary wider than one of RICOH Theta V, and the difference from the ground truth was large when I used the video captured with Insta360 One X2. Therefore, I would like to adjust parameters for Insta360 ONE X2.

Left: RICOH Theta V, Right: Insta360 ONE X2

I found the description of parameters (https://openvslam-community.readthedocs.io/en/latest/parameters.html), but I have no idea how to adjust these parameters especially ORB Parameters.

Could you tell me how to adjust parameters for the cameras not in example?

Config file for RICOH Theta V

# Equirectangular model for RICOH THETA S

#==============#
# Camera Model #
#==============#

Camera:
  name: "RICOH THETA V"
  setup: "monocular"
  model: "equirectangular"

  fps: 30.0
  cols: 3840
  rows: 1920

  color_order: "RGB"

#=====================#
# Tracking Parameters #
#=====================#

Tracking:
  max_num_keypoints: 5000

#================#
# ORB Parameters #
#================#

Feature:
  scale_factor: 1.2
  num_levels: 8
  ini_fast_threshold: 20
  min_fast_threshold: 7
  mask_rectangles:
    - [0.0, 1.0, 0.0, 0.1]
    - [0.0, 1.0, 0.84, 1.0]
    - [0.0, 0.2, 0.7, 1.0]
    - [0.8, 1.0, 0.7, 1.0]

#====================#
# Mapping Parameters #
#====================#

Mapping:
  baseline_dist_thr_ratio: 0.02

Config file for Insta360 ONE X2

# Equirectangular model for RICOH THETA S

#==============#
# Camera Model #
#==============#

Camera:
  name: "Insta360 One X2"
  setup: "monocular"
  model: "equirectangular"

  fps: 30.0
  cols: 5760
  rows: 2880

  color_order: "RGB"

#=====================#
# Tracking Parameters #
#=====================#

Tracking:
  max_num_keypoints: 5000

#================#
# ORB Parameters #
#================#

Feature:
  scale_factor: 1.2
  num_levels: 8
  ini_fast_threshold: 20
  min_fast_threshold: 7
  mask_rectangles:
    - [0.0, 1.0, 0.0, 0.1]
    - [0.0, 1.0, 0.84, 1.0]
    - [0.0, 0.2, 0.7, 1.0]
    - [0.8, 1.0, 0.7, 1.0]

#====================#
# Mapping Parameters #
#====================#

Mapping:
  baseline_dist_thr_ratio: 0.02

[ymd-stella]

I recommend that you try lowering the resolution by resizing.

[Shunmo17]

Thank you for your reply. I will try with 4K, or Full-HD resolution.

[Shunmo17]

I tried to resize the video to 4K, and I get better result.
The left figure is original 5.2K, and the right one is resized 4K.

Estimated trajectories in each trials (same video)

Error (Slow movement)

Error (Little fast movement)

Although the quality of estimated trajectories in each trial improved, compared to the result of ThetaV, the error to the ground truth is a little remained. What parameters should I adjust to further improve localization accuracy?

[ymd-stella]

If the other experimental conditions were the same, the only difference would have been in the camera used. If the error is greater when moving the camera faster, then I would be concerned about the automatic exposure adjustment and rolling shutter distortion. In addition, I think it is possible that the camera is doing some camera-specific image correction.

[ymd-stella]

I also think that increasing the frame rate will increase performance.

[Shunmo17]

Thank you for your advice. I will try high frame rate when moving fast.