For example, ./config/ifr-fusion-lr-kt0.yaml contains several settings:
dataset_type: "TUM" # for ICL-NUIM 's TUM format
scene: "0n" # scene 0 with noise
use_gt: False # not use groundtruth trajectory
pose_folder: "./treasure/orbslam2_record/lrkt0n/" # the predicted pose stream from orbslam2
outdir: "./res/lrkt0n_ours/" # where we stores the intermediary mesh
calib: [481.2, 480.0, 319.50, 239.50, 5000.0] # calibration of images
sequence_kwargs: # this is called with within sequence namespace
path: "./data/ICL_NUIM/lr_kt0n/"
start_frame: 0
end_frame: -1 # Run all frames
first_tq: [-1.4, 1.5, 1.5, 0.0, -1.0, 0.0, 0.0] # Starting pose
# Network parameters
training_hypers: "./treasure/hyper.json"
using_epoch: 600
# Enable visualization
vis: True
resolution: 4
# meshing
max_n_triangles: 4e6
max_std: 0.15 # 0.06
# These two define the range of depth observations to be cropped. Unit is meter.
depth_cut_min: 0.5
depth_cut_max: 5.0
# not exactly used in code, please follow real implementation
meshing_interval: 20
integrate_interval: 20
# Mapping parameters
mapping:
# Bound of the scene to be reconstructed
bound_min: [-3.5, -0.5, -2.5]
bound_max: [4.5, 3.5, 5.5]
voxel_size: 0.1
# Prune observations if detected as noise.
prune_min_vox_obs: 16
ignore_count_th: 16.0
encoder_count_th: 600.0