3dmatch.yaml

entangled: &entangled False

kpfcn_config:
  num_layers: 4
  in_points_dim: 3
  first_feats_dim: 256
  final_feats_dim: 32
  first_subsampling_dl: 0.025
  in_feats_dim: 1
  conv_radius: 2.5
  deform_radius: 5.0
  num_kernel_points: 15
  KP_extent: 2.0
  KP_influence: linear
  aggregation_mode: sum
  fixed_kernel_points: center
  use_batch_norm: True
  batch_norm_momentum: 0.02
  deformable: False
  modulated: False
  add_cross_score: True
  condition_feature: True
  coarse_feature_dim: &coarse_feature_dim 528
  fine_feature_dim:  264
  coarse_match_radius: 0.06
  coarse_level: -2


coarse_matching: &coarse_matching
  feature_dim: *coarse_feature_dim
  confidence_threshold: 0.2
  entangled: *entangled
  dsmax_temperature: 0.1
  match_type: &match_type 'dual_softmax' # [sinkhorn, dual_softmax]
  skh_init_bin_score: 1.0
  skh_iters: 3
  skh_prefilter: False


coarse_transformer:
  feature_dim: *coarse_feature_dim
  n_head: 4
  layer_types: ['self', 'cross', 'positioning', 'self', 'cross']
  positioning_type: &positioning_type 'procrustes' # options: [ 'randSO3' , 'procrustes', 'oracle']
  pe_type: &pe_type "rotary" # options: [ 'rotary', 'sinusoidal']
  vol_bnds: [ [ -3.6, -2.4,  1.14], [ 1.093, 0.78, 2.92 ]]
  voxel_size:  0.08 #default 0.08 # 1.6cm
  feature_matching : *coarse_matching
  entangled: *entangled
  procrustes:
    max_condition_num:  0 #30
    sample_rate: 1.0



train_loss:
  focal_alpha: 0.25
  focal_gamma: 2.0
  pos_weight: 1.0
  neg_weight: 1.0
  motion_loss_type: L1
  motion_weight: 0.
  match_weight: 1
  match_type: *match_type
  positioning_type: *positioning_type
  confidence_threshold_metric: 0.05
  mutual_nearest: False
  inlier_thr: 0.1 # 0.1m
  fmr_thr: 0.05
  registration_threshold: 0.2 # [3DMatch: 0.2m]
  dataset: 3dmatch



# misc
exp_dir:   test
mode: test
gpu_mode:  True
verbose: True
verbose_freq: 1000
do_valid: True
do_test: True
pretrain: 'pretrained/3dmatch/model_best_loss.pth'


# optimiser:
optimizer: SGD
max_epoch: 1500
lr: 0.015
weight_decay: 0.000001
momentum: 0.93
scheduler: ExpLR
scheduler_gamma: 0.95
scheduler_freq: 1
iter_size: 1



#dataset:
dataset: 3dmatch
folder: -last
batch_size: 1
num_workers: 8
augment_noise: 0.005
data_root: data/indoor
split: {
  "train": configs/indoor/train_info.pkl,
  "val": configs/indoor/val_info.pkl,
  "test": configs/indoor/3DMatch.pkl
#  "test": configs/indoor/3DLoMatch.pkl
}