-
Notifications
You must be signed in to change notification settings - Fork 1
/
hio_sdf_params.yaml
61 lines (54 loc) · 3.2 KB
/
hio_sdf_params.yaml
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
# Frames and topics
map_frame: 'map'
coarse_sdf_topic: '/voxblox_node/esdf_map_out'
pointcloud_topic: '/velodyne_points'
# Main options
device: 'cuda:0'
use_local_sdf_data: True # If True coarse SDF data is aggregated with local raycasting-based SDF data, otherwise only coarse SDF data is used
do_eval: False # Set to True if you want to run comparisons against the ground truth for a specific dataset (you will have to specify the "Evaluation parameters" below)
save_mesh: True # Saves the mesh of the environment (SDF zero-level set) after each training iteration in the output folder
save_slices: False # Saves slices of the SDF at pre-defined height intervals after each training iteration in the output folder
publish_local_data: True # Publishes the current local point cloud-based 3D samples
publish_sdf_data: False # Publishes a current slice from aggregated (discrete) SDF data at height sdf_slice_z
publish_sdf_slices: True # Publishes current SDF slices from the (continuous) trained model at resolution visualization_resolution
sdf_slice_z: 0.5 # Desired height of visualized coarse SDF slice
sdf_visualization_resolution: 0.05 # Desired resolution for SDF slice visualization of trained model
stream_weights_over_zmq: False # (Optional) Sets up a ZMQ server that can stream the most recent HIO-SDF weights
# Voxfield parameters
voxel_size: 0.1 # IMPORTANT: Make sure this matches your Voxfield configuration
voxel_per_side: 8 # IMPORTANT: Make sure this matches your Voxfield configuration
kEpsilon: 0.000001
kNumDataPacketsPerVoxel: 2
# Voxfield sampling parameters
num_voxfield_samples_surface: 10000 # Max number of used Voxfield samples lying close to the surface for each training iteration
num_voxfield_samples_freespace: 30000 # Max number of used Voxfield samples away from the surface for each training iteration
# Parameters for local SDF sampling
num_local_pcds: 1 # (Optional) If >1 a window of several local point cloud measurements are aggregated to a single local point cloud
num_local_pcd_points: 50000 # Max number of points in local point cloud (point cloud is sub-sampled to this number)
max_raycasting_distance: 6.0 # Maximum raycasting distance for local SDF samples
truncation_distance: 0.2 # Truncation distance behind the observed point cloud for local SDF samples
num_raycasting_points: 20 # Max number of sampled points along each ray
num_local_samples: 20000 # Max number of local SDF samples for each training iteration
# Training parameters
learning_rate: 0.0004
weight_decay: 0.012
sdf_loss_weight: 5.0
eikonal_loss_weight: 2.0
warm_start_iters: 5
epochs_warm_start: 50
epochs_nominal: 10
# Evaluation parameters
gt_sdf_file: '/home/vasileiosv/gt_sdfs/apt_3/1cm/sdf.npy'
sdf_transf_file: '/home/vasileiosv/gt_sdfs/apt_3/1cm/transform.txt'
dataset_format: 'ReplicaCAD' # other formats are 'ScanNet' and 'SAIC-NY'
scene: 'apt_3_nav'
grid_resolution: 0.05 # Resolution of global regular grid used for evaluation
eval_point_samples: 25000 # Maximum number of samples from the regular used for evaluation
regular_grid_x_min: -20.0
regular_grid_x_max: 20.0
regular_grid_y_min: -20.0
regular_grid_y_max: 20.0
regular_grid_z_min: -1.0
regular_grid_z_max: 4.0
# Output path
output_path: '/home/vasileiosv/global_sdf_results/hio_sdf/'