Skip to content

Latest commit

 

History

History
42 lines (32 loc) · 2.25 KB

README.md

File metadata and controls

42 lines (32 loc) · 2.25 KB

JSON configuration parameters

Precomputation

  • pre_min_cube_size: Voxel size at the lowest resultion. It is doubled at each downsampling stage.

  • pre_filter_size: Number of elements in each dimension of the convolutional kernel.

  • pre_num_scales: The number of resolutions to use.

  • pre_num_rotations: Number of rotations used for rotational augmentation.

  • pre_num_neighbors: Number of neighboring points used to precompute the structure of conv kernel.

  • pre_noise_level: Standard deviation of the additive Gaussian noise on point locations.

  • pre_output_dir: Directory for storing data after pre-computation.

  • pre_interp_method: Method for interpolating the signal in tangent images ('depth_densify_nearest_neighbor', 'depth_densify_gaussian_kernel').

  • pre_dataset_param: Dataset type ('stanford', 'scannet' or 'semantic3d').

Common

  • co_train_file: List of training scans.
  • co_test_file: List of test scans.
  • co_experiment_dir: Path to current experiment directory.
  • co_output_dir: Relative path to store network outputs.

Training and testing

  • tt_log_dir: Directory where to output logs.
  • tt_snapshot_dir: Directory for saving network snapshots.
  • tt_input_type: Which input features to use for training. A string containing one or more of the following values: c (color), d (depth), n (normals), h (height).
  • tt_max_snapshots: Maximum number of snapshots to be saved.
  • tt_test_iter: Frequency of running the validation during training (in iterations).
  • tt_reload_iter: Frequrency of updating the training set (in iterations). Used because of the rotational augmentation, to not store all scans in memory all the time.
  • tt_max_iter_count: Maximum number of training iterations.
  • tt_batch_size: Batch size.
  • tt_valid_rad: Radius of a sphere to be used for sampling when working with large scans.
  • tt_filter_size: Size of the convolutional filter along one dimension.
  • tt_batch_array_size: Number of batches to pre-load when sampling from large scans.

Evaluation

  • eval_scan_file: Name pattern for the raw scan file.
  • eval_label_file: Name pattern for the raw label file.
  • eval_output_file: Name pattern for file with extrapolated labels.