hello,@LimHyungTae ，When performing matching in long corridors, using Quatro as the initial value for solution calculation revealed that using Quatro worsened the results, leading to a greater need for correction in later ICP calculations.
The project I'm using is fast_lio_localization_sc_qn, but after successful relocalization, I've been using Quatro+ICP matching. The image process is as follows: first, fastlio is used to obtain the odometry pose; then Quatro provides a coarse solution; and finally, ICP provides a fine solution.
config:
quatro: # all values are from Quatro official github
enable: true # whether or not to use quatro, if false -> FAST-LIO-SAM-N (Nano-GICP only)
optimize_matching: true # whether or not to use Optimized matching
use_submap_accum: true # true: Quatro 使用子图累积目标地图；false: 仅使用单关键帧
distance_threshold: 30.0 # when optimized matching, use only correspondences within this radius, unit [meter]
max_correspondences: 350 # max correspondences to use for optimized matching
fpfh_normal_radius: 0.6 # should satisfy the following condition: fpfh_radius >= 1.5 * fpfh_normal_radius
fpfh_radius: 0.9 # should satisfy the following condition: fpfh_radius >= 1.5 * fpfh_normal_radius
estimating_scale: false
noise_bound: 0.3 # The magnitude of uncertainty of measurements, the best is within v/2 ~ v (v: voxel resol.)
rotation:
num_max_iter: 50 # Usually, rotation estimation converges within < 20 iterations
gnc_factor: 1.4 # Control the magnitue of the increase in non-linearity. The larger the value, the steeper the increase in nonlinearity.
rot_cost_diff_threshold: 0.0001 # The cos
hello,@LimHyungTae ，When performing matching in long corridors, using Quatro as the initial value for solution calculation revealed that using Quatro worsened the results, leading to a greater need for correction in later ICP calculations.
The project I'm using is fast_lio_localization_sc_qn, but after successful relocalization, I've been using Quatro+ICP matching. The image process is as follows: first, fastlio is used to obtain the odometry pose; then Quatro provides a coarse solution; and finally, ICP provides a fine solution.
config:
quatro: # all values are from Quatro official github
enable: true # whether or not to use quatro, if false -> FAST-LIO-SAM-N (Nano-GICP only)
optimize_matching: true # whether or not to use Optimized matching
use_submap_accum: true # true: Quatro 使用子图累积目标地图；false: 仅使用单关键帧
distance_threshold: 30.0 # when optimized matching, use only correspondences within this radius, unit [meter]
max_correspondences: 350 # max correspondences to use for optimized matching
fpfh_normal_radius: 0.6 # should satisfy the following condition: fpfh_radius >= 1.5 * fpfh_normal_radius
fpfh_radius: 0.9 # should satisfy the following condition: fpfh_radius >= 1.5 * fpfh_normal_radius
estimating_scale: false
noise_bound: 0.3 # The magnitude of uncertainty of measurements, the best is within v/2 ~ v (v: voxel resol.)
rotation:
num_max_iter: 50 # Usually, rotation estimation converges within < 20 iterations
gnc_factor: 1.4 # Control the magnitue of the increase in non-linearity. The larger the value, the steeper the increase in nonlinearity.
rot_cost_diff_threshold: 0.0001 # The cos