Skip to content

Latest commit

 

History

History
109 lines (95 loc) · 5.47 KB

configs.md

File metadata and controls

109 lines (95 loc) · 5.47 KB
Raw

Description of configs.yaml

1. Example Configuration Files:

We provide some example configuration files in ../configs/. They are for the different settings in the paper, including:

Configuration Name Experiment Setting
icp.yaml ICP + Motion Prior
icp_s1.yaml ICP + Motion Prior + Shape Term (First Frame as Shape)
icp_sall.yaml ICP + Motion Prior + Shape Term (Aggregated Shape)
icp_sall_m.yaml ICP + Motion Prior + Shape Term (Aggregated Shape) + Motion Consistency

2. Description on Reading / Writing Configuration Files

The fields and the description of configs.yaml are as follows. Please look at the corresponding components to see how we may initialize the whole tracker using the configurations.

running:                     # running configurations
  window_size: 2             # how many frames to consider 
    
data_loader:                 # data loader configurations
  terrain: false             # whether to load the terrain map
    
finder:                      # configurations for finding the motion between frame 0-1
  optim:                     # optimizer configurations
    method: BFGS             # optimizer type
    ITER: 50                 # iterations for iterations out-of-optimizer
    options:                 # optimizer related options
      maxiter: 50            # maximum iterations for optimizer
      disp: false    
  neighbor_num: 1            # nearest neighbor number for icp loss computation
  box_scaling_prev: 1.0      # when segmenting the related point cloud, the scale in the previous box    
  box_scaling_next: 3.0      # when segmenting the related point cloud, the scale in the next box    
  least_pc: 10               # least LiDAR point number to consider   
  pc_limit: 10000            # maximum LiDAR point number, downsample the number of surpass the limit    
  loss_type: L2              # loss type, such as L2, Huber
  ransac:                    # ransac options
    switch: false            # if use ransac
    
optim:                       # configurations for the optimizer in tracking
  method: BFGS               # optimizer type
  iter: 20                   # iteration out-og-optimizer
  options:                   # optimizer options
    maxiter: 50              # maximum optimizer iterations
    disp: false    
    
motion_model:                # motion model related configurations
  moving_avg_weight: 0.5     # moving average weight
    
shape_map:                   # configurations for maintaining a shape during tracking
  update_freq: 1000          # how many frames for an update, this 1000 means no update at all on waymo
  subshape_len: 2            # subshape length
  box_scaling_pc_bank: 1.0   # the scale for segmenting the related LiDAR points
  downsample: true           # if downsample the shape
  resolution: 0.05           # the voxel size for downsampling

factors:                     # loss factors
  switch:                    # the switch for some factors, might save time
    motion_prior: true       # whether to use the motion_prior term
    icp_loss: true
    shape_loss: true
    motion_consistency: true
    latitude: false
    detection: false         
  names:                     # factor names
    - latitude
    - icp_loss
    - shape_loss
    - motion_prior
    - motion_smooth
    - motion_consistency
  latitude:                  # latitude factor configurations
    scaling: 1.0             # when get the latitude, how much terrain point cloud to look at
  icp_loss:                  # icp loss between neighboring frames
    neighbor_num: 1          # nearest neighbor number
    least_pc: 10             # least LiDAR point number to consider  
    pc_limit: 1000           # maximum LiDAR point number, downsample the number of surpass the limit   
    box_scaling_prev: 1.1    # when segmenting the related point cloud, the scale in the previous box
    box_scaling_next: 1.5    # when segmenting the related point cloud, the scale in the next box    
    loss_type: L2            # loss type, such as L2, Huber
    ransac:                  # ransac options
      switch: false          # if use ransac
      num_iter: 200          # ransac iterations
      threshold: 0.2         # ransac outlier distance threshold
      ransac_limit: 10       # least inlier / LiDAR point number to consider
  shape_loss:                # shape loss configurations
    neighbor_num: 1          # nearest neighbor number
    least_pc: 10             # least LiDAR point number to consider  
    pc_limit: 1000           # maximum LiDAR point number, downsample the number of surpass the limit   
    box_scaling_next: 1.5    # when segmenting the related point cloud, the scale in the next box    
    loss_type: L2            # loss type, such as L2, Huber
    ransac:                  # ransac options
      switch: false          # if use ransac
      num_iter: 200          # ransac iterations
      threshold: 0.2         # ransac outlier distance threshold
      ransac_limit: 10       # least inlier / LiDAR point number to consider
  motion_prior: ~            # motion prior factor configurations
  motion_consistency: ~      # motion model loss configurations

weight:                      # loss weights
  icp: 1.0
  latitude: 1.0
  motion_prior: 0.1
  motion_smooth: 0.0
  shape_loss: 1.0
  motion_consistency: 0.0